The Economics of Hedge Fund Startups: Theory and Empirical

THE JOURNAL OF FINANCE • VOL. , NO. 0 • FEBRUARY 2021

The Economics of Hedge Fund Startups: Theoryand Empirical Evidence

CHARLES CAO, GRANT FARNSWORTH, and HONG ZHANG

ABSTRACT

This paper examines how market frictions influence the managerial incentivesand organizational structure of new hedge funds. We develop a stylized model inwhich new managers search for accredited investors and have stronger incentivesto acquire managerial skill when encountering low investor demand. Fund familiesendogenously arise to mitigate frictions and weaken the performance incentivesof affiliated new funds. Empirically, based on a TASS-HFR-BarclayHedge mergeddatabase, we find that ex ante identified cold inceptions facing low investor demandoutperform existing hedge funds and hot inceptions facing high demand and that coldstand-alone inceptions outperform all types of family-affiliated inceptions.

THE HEDGE FUND INDUSTRY HAS EXPERIENCED dramatic growth over thelast few decades. For example, worth less than $100 billion prior to the1990s, it ballooned to $3 trillion in assets under management (AUM) by2019. Although capital flows to both existing and new funds are importantin explaining the rapid growth of the hedge fund industry, the literature hasfocused primarily on the former. The pioneering work of Aggarwal and Jorion(2010) on hedge fund inceptions presents evidence of outperformance duringthe first two or three years of existence. They also find strong evidence thatearly performance by individual hedge funds is persistent. Since a competitive,frictionless market allows capital to flow freely across fund types and receive

Charles Cao is at Pennsylvania State University. Grant Farnsworth is at Texas Christian Uni-versity. Hong Zhang is at Tsinghua University. We are grateful to Stefan Nagel (the Editor), theAssociate Editor, and two anonymous referees for many insightful suggestions that have greatlyimproved the quality of the paper. We also thank Vikas Agarwal; Amber Anand; Chris Ander-son; Adrien Becam; Utpal Bhattacharya; John Bizjack; Yong Chen; Serge Darolles; Robert DeY-oung; Darrell Duffie; Heber Farnsworth; Wayne Fearson; Nicolae Gârleanu; Bing Han; Sara Hol-land; Paul Irvine; Shane Johnson; Chan Kim; Paul Koch; William Kracaw; Bing Liang; AndrewLo; David Rapach; Mo Rodriguez; Ravi Shukla; Tim Simin; Matthew Spiegel; Kjetil Storesletten;Melvyn Teo; Jules van Binsbergen; David Weinbaum; and seminar participants at PennsylvaniaState University, the 2015 Society of Financial Studies (SFS) Finance Cavalcade, the 2018 Amer-ican Finance Association (AFA) Meeting, and the 9th Annual Hedge Fund and Private EquityResearch Conference for helpful comments. Research assistance by Han Xiao is greatly appreci-ated. The authors do not have any potential conflicts of interest to disclose, as identified in TheJournal of Finance disclosure policy.

Correspondence: Charles Cao, 338 Business Building, University Park, PA 16802, USA. 1-814-865-7891; e-mail: [email protected].

DOI: 10.1111/jofi.13009

© 2021 the American Finance Association

1

mailto:[email protected]

2 The Journal of Finance®

comparable risk-adjusted returns (see Berk and Green (2004), Berk and vanBinsbergen (2015, 2017)), these findings raise important questions: do marketfrictions hinder the flow of capital in the hedge fund industry and, if so, dothese frictions shape the managerial incentives and organizational structuresof the hedge fund industry.

In this paper, we address these questions by analyzing one of the most im-portant types of market frictions faced by new hedge fund managers, namely,the need to search for accredited investors. Prior work shows that search fric-tions are important for mutual funds (see Sirri and Tufano (1998), Choi, Laib-son, and Madrian (2010) for empirical evidence and Hortaçsu and Syverson(2004), Gârleanu and Pedersen (2018) for theoretical treatments). Becausehedge funds face many marketing restrictions, search frictions may play aneven more important role for these funds. However, the burden of search forhedge funds differs from that of mutual funds: instead of investors using pub-lic information to search, as discussed in mutual fund studies, new hedge fundmanagers often must find accredited investors and persuade them to invest.

We develop a stylized model in which we incorporate into the model of Berkand Green (2004) the need for a new manager to raise capital. In the spirit ofRubinstein and Wolinsky (1985) and Duffie, Gârleanu, and Pedersen (2005,2007), we model fund-raising as a two-step search-and-bargaining process.This framework delivers several novel predictions and sheds light on the crit-ical role played by search frictions in the hedge fund industry. The need tosearch for investors not only influences managerial incentives of new fundsbut also drives the formation of fund families.

To see how search frictions influence managerial incentives, we note thatthe total capital raised is determined by two margins: the extensive margin,which refers to the investors and initial capital that the manager identifiesvia search, and the intensive margin, the fraction of capital that the managerretains after bargaining with the investors. While the extensive margin is re-lated primarily to investor demand, the intensive margin can be influenced bythe merit of the fund: by investing in costly credible skills necessary to deliversuperior (expected) performance, the manager can persuade a larger fractionof matched investors to contribute.

Importantly, the two margins act as substitutes, which impacts managerialincentives. A high extensive margin reduces a manager’s incentive to use su-perior performance to persuade investors in the bargaining step. A novel andtestable implication arises. If we refer to new funds launched using a “hot”strategy (i.e., a hedge fund strategy that is popular among investors at thetime) as hot inceptions and those launched using a “cold” strategy (an unpopu-lar strategy) as cold inceptions, our model predicts that cold inceptions shouldoutperform hot inceptions.

Search frictions also provide an economic rationale for a key organizationalfeature of the industry: hedge fund families. Two types of family-affiliated in-ceptions arise in our model. First, family structures emerge to allow affiliatednew funds to benefit from existing funds’ investor pool. These investors mayinvest in the new fund or may introduce other accredited investors to the fund.

The Economics of Hedge Fund Startups 3

The latter networking effect reduces new funds’ search costs and is consistentwith the effects of social networks (e.g., Jackson and Rogers (2007)). However,the search advantage provided by family structures reduces the performanceincentive of affiliated new managers. Our model therefore predicts that family-affiliated inceptions deliver poorer performance than stand-alone inceptions.

The second channel giving rise to family-affiliated inceptions is diseconomiesof scale, a widely observed feature of hedge funds. In our model, search frictionsamplify diseconomies of scale. Thus, when investor demand for existing fundsexperiences a positive shock, fund families have incentives to launch clone in-ceptions that closely mimic existing funds to absorb the excess demand. Cloneinceptions are de facto hot and unlikely to contribute new skills and deliversuperior performance.

We test these model predictions using a comprehensive sample of hedgefunds that we obtain by merging three leading commercial hedge funddatabases—Lipper TASS, HFR, and BarclayHedge—over the period 1994 to2016. Specifically, we conduct three tests to investigate the performance differ-ence between cold and hot inceptions.

First, we exploit variation in the popularity of hedge fund strategies amonginvestors. Since investors chase past performance, we use recent strategy re-turns and flows to capture strategy popularity. Empirically, we find that coldinceptions deliver better performance than both existing funds and hot incep-tions. Over the 60-month holding period after initial inception, cold inceptionsoutperform hot inceptions by 0.24% per month (or 3% annually) on a risk-adjusted basis.

Second, we explore the role of hedge fund families. We find that stand-aloneinceptions outperform family-affiliated inceptions by 0.23% per month (or 2.8%per year) on a risk-adjusted basis. For family-affiliated nonclone funds, coldinceptions outperform hot ones by 4.3% annually. In contrast, we do not find aperformance difference between cold and hot clone inceptions.

The above results suggest that an empirical strategy combining cold stand-alone and hot clone inceptions will have the most power to identify the effects ofextensive-margin advantages because this strategy incorporates the influencesof both strategy demand and family structure. The performance gap betweencold stand-alone inceptions and hot clone inceptions is as high as 0.55% permonth (or 6.8% annually), which is statistically and economically significant.Due to its appealing economic interpretation, we adopt this empirical strategyin several tests below. The results provide strong support to the predictionthat superior-performing new hedge funds can be identified ex ante based onan understanding of the effects of investor demand and family structures.

Since our model applies best to new fund managers (experienced managersmay have access to more investors), we also conduct a test focusing on theinceptions of new managers. To do so, we exclude inceptions run by managerswho have previously managed other funds. The results using this sample aresimilar to those of our main tests but have larger economic magnitudes.

We next examine the economic source of cold inceptions’ outperformance. Inour model, superior performance is driven by investment in managerial skills.


However, a leading concern about hedge fund performance is that greaterexposure to illiquidity or deliberate return-smoothing may allow some fundsto inflate performance (e.g., Getmansky, Lo, and Makarov (2004)). We conducta battery of tests to investigate the source of the performance difference be-tween cold and hot inceptions and the extent to which it reflects genuine skill.

We first examine the security-selection and market-timing skills (e.g.,Treynor and Mazuy (1966)) of cold and hot inceptions. We find that managersof cold inceptions exhibit significant skill in security-selection but no skill inmarket-timing. In contrast, managers of hot inceptions demonstrate negative(incorrect) market-timing ability and weaker selection skill, with the net effectthat they do not deliver alpha.

Next, since persistence analysis provides a powerful test for managerial skill(e.g., Carhart (1997)), we examine whether there is any difference in perfor-mance persistence between cold and hot inceptions. We find that the perfor-mance of cold inceptions is highly and significantly persistent over time, whilehot inceptions exhibit negative or insignificant persistence.

Finally, we show that the performance gap between hot and cold inceptionscannot be attributed to illiquidity or return-smoothing. Additionally, it cannotbe explained by risk factors beyond the Fung and Hsieh (2004) seven factors, byfund characteristics, or by fund policy choices. Rather, our subsample analysisof convertible arbitrage (CA) funds suggests that cold inceptions exploit moresophisticated economic sources than market-wide risk or well-known arbitrageopportunities.

Backfill bias is an important concern in hedge fund research. It arises whenmanagers joining a database have the option of reporting performance betweeninception and the initial report date. Because funds may not report if early per-formance is poor, reported returns from the backfill period exhibit an upwardbias. To account for this bias, in all tests we use the approach of Jorion andSchwarz (2019) to identify the add dates and we delete all observations beforethis date.

This study builds on and extends the work of Aggarwal and Jorion (2010),who use a novel event-time approach and careful controls for backfill bias toshow that new funds deliver alpha and performance persistence during thefirst two to three years. Their findings highlight the importance of new tal-ent entering the industry but also strongly suggest that market frictions ex-ist that hinder the efficiency of capital flows to new funds. Sun, Sun, andZheng (2020) study whether investor sentiment affects the decision to startnew funds and document a significant positive impact. We develop a model toexplore these issues and show that search frictions affect portfolio manage-ment through family structure and negative demand-performance incentives.Both mechanisms are novel to the literature and play an important role indetermining the cross-section of new hedge fund performance. Unlike in Ag-garwal and Jorion (2010), these mechanisms lead us to explore and documentsignificant performance heterogeneity across various types of hot and coldinceptions.

An emerging literature examines the influence of market frictions, par-ticularly search frictions, on delegated portfolio management. Theoretical


treatments of search frictions concentrate on the costs that investors bear insearching for funds (e.g., Hortaçsu and Syverson (2004), Gârleanu and Peder-sen (2018)).1 We complement these studies by examining the influence of newhedge fund managers’ need to search for accredited investors. Building on thework of Berk and Green (2004) and the search framework of Rubinstein andWolinsky (1985) and Duffie, Gârleanu, and Pedersen (2005, 2007), we showthat this friction critically influences the incentives of new managers and theorganizational structure of the industry. Our model is tractable and its predic-tions are consistent with the data.

Finally, this study extends the literature on hedge fund and mutual fundfamilies. Although family structures are widely observed in both hedge fundand mutual fund industries, the underlying economic rationales differ betweenthese industries. Our model suggests that hedge fund families can arise tomitigate search frictions or to address search-enhanced diseconomies of scale.The predictions of the model are supported in our empirical findings.

The remainder of the paper proceeds as follows. Section I presents a modelof hedge fund inceptions and its testable hypotheses. Section II describes thehedge fund data that we use in our analysis. Section III examines the determi-nants of hedge fund inception probability. Section IV studies the influence ofstrategy demand and family structure on the performance of inceptions. Sec-tion V explores alternative explanations for our findings. Finally, Section VIconcludes.

I. Theoretical Framework for Hedge Fund Inceptions

In this section, we develop a stylized model of hedge fund startups. We ex-tend the model of Berk and Green (2004; hereafter, BG) by incorporating a keyfeature that affects hedge fund inceptions, namely, managers’ need to searchfor accredited investors.2

A. The BG Benchmark of Existing Funds in the Same Strategy Category

Before we examine the launch of a new fund in a given strategy, we describeexisting funds in the same strategy category. For tractability, existing fundsin the strategy are represented by a benchmark fund whose operation and

1 Hortaçsu and Syverson (2004) show that investors’ search costs help explain the puzzling feedispersion among S&P 500 index funds. Gârleanu and Pedersen (2018) examine the asset pricingimplications of an extended Grossman-Stiglitz (1980) model in which investors search for mutualfunds. In addition to search frictions, Jylhä and Suominen (2011) show that, in a two-countrymodel, hedge funds arise endogenously to mitigate market segmentation, while Glode and Green(2011) model the bargaining process between hedge fund managers and investors.

2 Indeed, raising money is widely regarded as one of the most difficult tasks of a new hedgefund (see, e.g., the discussion on how to start a hedge fund at https://www.lifeonthebuyside.com/start-a-hedge-fund/). In practice, new managers often need to actively search for potential in-vestors, for instance, through networking. Even when potential accredited investors are found,it is not an easy task to raise capital, as many such investors have professional teams to aid in-vestment (e.g., the family office of wealthy families).

https://www.lifeonthebuyside.com/start-a-hedge-fund/

https://www.lifeonthebuyside.com/start-a-hedge-fund/


dynamics follow BG, except for an additional search term that we specify below.Although the BG model was originally designed for mutual funds, its two keyfeatures—diseconomies of scale in fund operation and the equilibrium of theindustry achieved through fund size—apply well to the hedge fund industry.In particular, as Berk and van Binsbergen (2015, 2017) point out, the marketfor mutual funds equilibrates in quantity as the price for funds is fixed by themarket value of the funds’ underlying assets. In that context, fund size proxiesfor managerial skill. In our model, the hedge fund industry achieves a similarequilibrium with an additional key influence—search frictions.

Following the notation of BG, we assume that in a given investment period t,a benchmark hedge fund is endowed with the skill to generate a risk-adjustedstrategy benchmark return of Rt,with expected value φt−1,which is observableto investors.3 Further, we assume that the fund distributes a cost-adjusted re-turn of rt = Rt − c(qE

t ) − s(qEt ), where qE

t denotes fund size of existing funds,E. The variable c(qE

t ) = b × qEt + f is the fund-size-normalized cost function

following Berk and van Binsbergen (2015, 2017), where b and f denote opera-tional costs (with diseconomies of scale) and management fees.4 The last term,s(qE

t ), we introduce in the model to describe the size-normalized search costthat the fund incurs to raise capital qE

t . The search cost can be thought of asa networking and marketing cost that can be deducted from the payoff of thefund (we specify the cost below).

When investors receive zero net-of-fee returns, we have the following search-enhanced BG equilibrium condition (hereafter, the BG condition):

E (rt+1) = φt − c(qE

t

) − s(qE

t

) = 0. (1)

This condition says that the fund industry equilibrates in fund size and thatthe fund manager earns the economic rents that she creates. We assume thissplit of the benchmark fund’s economic rents to highlight new funds’ incen-tive problem. Adjusting the split will not affect the incentive difference acrossdifferent types of funds.

To model managerial incentives, we follow BG and assume that a fund man-ager benefits from more capital: she derives a utility gain of g(qt ) = f × q(t) bymanaging a fund of size q(t). Here, we remove the superscript E because theutility applies to both new and existing funds. At the same time, the managercan enhance a fund’s expected risk-adjusted return by δ if she pays a linear

3 More explicitly, the fund can generate a risk-adjusted return of Rt = α + εt , where α ∼N(φ0, η

2), based on known information, denotes the performance of the strategy (i.e., risk-adjustedreturn funds in this strategy category deliver) and εt ∼ N(0, σ 2) is noise. Investors do not observethe true distribution of managerial skill. Rather, they use the realized return to update their prior,and expect that the benchmark fund will deliver an expected return of φt−1 ≡ E(Rt |Rt−1, . . . , R1)in period t.

4 BG assumes that the dollar cost of operation exhibits diseconomies of scale; c(qEt ) is the dollar

cost scaled by fund size.


private learning cost, L(δ) = L0 × δ, where L0 is a positive coefficient.5 Themanager can then deliver δ to investors to attract more capital.6

The trade-off between the marginal benefits and marginal costs determinesthe optimal level of performance that the manager wants to deliver. Althoughthis trade-off resembles BG, the critical difference is that fund managers alsoface search frictions, which we examine next.7

B. Capital-Raising as a Search-and-Bargaining Process

Search can go in two directions: investors can search for managers, and man-agers can search for investors. A close look at an investor-search BG modeland its comparison to the data (see Section II of the Internet Appendix) sug-gests that new hedge funds may benefit from actively reaching out to investors,giving rise to manager-initiated search. We therefore adopt the search frame-work of Rubinstein and Wolinsky (1985) and Duffie, Gârleanu, and Pedersen(2005, 2007) to examine how manager-initiated search affects the incentivesof new funds.8 To do so, we introduce three sets of assumptions that help ex-tend BG under these frictions: investor heterogeneity in supplying capital tohedge funds, capital-raising as a search-and-bargaining process, and familyaffiliation as a source of capital.

We first describe investor heterogeneity and capital. For a given category,its full set of existing and potential investors can be classified into four types,denoted by � = {ho, hn, lo, ln}, based on two sets of characteristics: “h” (high)and “l” (low) refer to an investor’s intrinsic preference for the hedge fund strat-egy (i.e., investors with a high (low) preference are willing (unwilling) to investin hedge funds in that strategy), while “o” and “n” refer to “old” investors ofthe existing fund and noninvestors who have not invested. For instance, “hn”refers to investors who have not invested in the existing fund of the strategybut are willing to do so if a fund in that strategy solicits capital from them.

We normalize the total mass of all investors to one (there is a continuumof investors) and assume that investors carry with them an amount of capi-tal to invest in period t, z(t). The variable z(t) describes investors’ aggregatedemand for the strategy, which is exogenous to the new fund by assumption.Hence, if we denote the fraction of type-σ investors (σ ∈ �) by μσ (t), such that

5 Although learning costs are often assumed to be convex, this assumption is not necessary inour model because the benefit of learning is concave in alpha. Having convex learning costs doesnot affect our main conclusions, as we will see below.

6 The fund can of course distribute only a fraction of alpha to investors and use the remainingalpha either to enlarge fund size as in the BG condition or it can retain the remaining alpha asincentive fees. The Internet Appendix shows that our main predictions remain valid in these cases.The Internet Appendix is available in the online version of this article on The Journal of Financewebsite.

7 Another difference is that hedge funds can use leverage, which we do not explicitly examinein this paper. Hence, we can interpret qt as leverage-adjusted fund size when hedge funds havealready taken the maximum amount of leverage. We thank the Associate Editor for this intuition.

8 Duffie (2010) provides more discussions on the mechanisms of imperfect move of capital.


μho(t) + μhn(t) + μlo(t) + μln (t) = 1, we can interpret μσ (t) as the mass of typeσ investors, who carry with them μσ (t) × z(t) of capital.

Next, we assume that capital-raising occurs at the beginning of period t,during which the existing fund and a new fund sequentially raise capital. Theexisting fund moves first. Once its steady state (characterized by stabilizedμσ (t)) is achieved, the fund collects capital from its investors (i.e., of type σ ∈{ho, lo}).9 The new manager then seeks to raise capital from the remaininginvestors in the market (i.e., of type σ ∈ {hn, ln}). Both funds invest the capitalraised during the period and deliver cost-adjusted payoffs to investors at theend of the period.

We explicitly model capital-raising as a search-and-bargaining procedure.Take the new fund as an example. In the search step, the new manager triesto find investors in the market. We assume that by paying a total searchcost of TS, a new fund can be matched with investors with an intensity ofρN, where the superscript N denotes new funds, and paying a higher searchcost allows the manager to be matched with a higher intensity of investors,that is, ρN = a × TS, where a is a positive constant. The extensive margin ofcapital-raising is then the total amount of matched capital in the search step,ρN (μhn(t) + μln(t))z(t).

Next the new manager bargains with investors, trying to persuade themto invest. Of matched investors, the hn-type will immediately invest due totheir high intrinsic preference, but ln-type investors are unwilling to do so.However, the manager can use her fund-specific performance, δ, to bargain.Since ln-investors receive an expected abnormal return of δ, traditional finan-cial theories (e.g., the CAPM) suggest that their optimal asset allocation in thenew fund would be in proportion to δ. Hence, without loss of generality, we as-sume that ln-investors can be persuaded with probability ξ (δ) = ξ0 × δ, ξ0 > 0,which allows the manager to raise capital in the amount of ρNμln(t)ξ (δ)z(t)from ln-investors.10 Superior performance, therefore, increases the fraction ofcapital that the manager can retain under the intensive margin of the capital-raising process.

9 Note that the value of μσ (t) is determined in the steady-state by the expected performanceφt of the existing fund, the search cost it pays, and the probabilities of h- and l-type investorsswitching preferences—we provide details on the search process and the steady-state of the oldfund in Lemma IA.2 in the Internet Appendix. For now, we note that investors remain in themarket from whom the new fund can raise capital.

10 There is no information asymmetry about δ in our model. It is worth noting that the secondstep in our model is equivalent to the bargaining process of a typical search-based asset pricingmodel in determining the equilibrium conditions. The difference is that in asset pricing modelsinvestors typically bargain for price (e.g., Duffie, Gârleanu, and Pedersen (2005, 2007)), whilein our model the manager bargains for the quantity of capital to be invested when the price isfixed. This notion of bargaining (for capital) is consistent with both the BG equilibrium conceptand the take-it-or-leave-it Nash bargaining models widely used in the search literature (see, e.g.,McMillan and Rothschild (1994) for a survey on such models). The Internet Appendix (at the endof the proof of Lemma 3) sketches a general search framework that generates the assumptionof having two types of investors with heterogeneous intrinsic preferences and the assumption ofincreasing bargaining power based on superior performance.


In economic terms, the search and bargaining steps capture the extensiveand intensive margins of the capital-raising process, during which a total ofρN (μhn(t) + μln(t)ξ (δ))z(t) can be raised. When the amount of capital is greaterthan or equal to the optimal fund size, q(t), the manager stops searching andlocks in the capital for investment:

ρN (μhn (t) + μln (t) ξ (δ)) z (t) ≥ q (t) . (2)

The capital-raising process above applies to both new and existing funds.Notably, the existing fund has an advantage of existing investors for which itssearch cost is zero. Thus, it searches for new investors as a replacement forwithdrawals by existing investors. In contrast, the burden of new funds is tofind all new investors. When the search process is costly, this difference givesrise to performance heterogeneity across funds.

If the existing fund benefits from its existing investors, a family structurecan endogenously arise to allow some new funds to benefit from these investorsas well. To capture this effect, we assume that new hedge funds can be eitherstand-alone or affiliated with the existing fund through a family structure, andwe separately examine the incentives and performance of these two types ofnew funds. We assume that a family structure enhances the mass of high-typeinvestors available to its affiliated new managers, that is, μN′

hn (t) = μhn(t) +γhμ

Eho(t), where μE

ho(t) is the high-type investors of the existing fund and γh(> 0)describes the networking effect of the existing investors in supplying capital,referring new investors, and providing relevant information.11

Conditioning on the additional assumptions above, a fund manager maxi-mizes her utility by optimizing search cost and fund-specific performance asfollows:

maxs(t), δ

U (s (t) , δ) = g (qt ) − L (δ) (3)

s.t. (1) and (2).

C. The Incentives of Stand-Alone New Funds in Generating Fund-SpecificAlpha

We first derive a closed-form solution to the managerial problem of stand-alone inception. Easy access to capital during the search step of the fund-raising process (a high extensive margin) influences the incentive to generateextra performance. The extensive margin is associated with investor demand,z(t), which is exogenous to the new fund. The impact of demand on managerialincentives is as follows.

11 Existing investors may also help managers reduce the intensity of low-type investors, that is,μN′

ln (t) = μNln(t) − γlμ

Elo(t), where γl is a positive parameter. Adding this effect does not affect our

main conclusions.


PROPOSITION 1: When a new manager solves the general problem (3) subjectto the conditions specified in (1) and (2), the following properties hold:

(i) The optimal level of fund-specific alpha is δ∗ =Max{0, f

12 (L0bμln(t)ξ0azt )

− 12 − μhn(t)

μln(t)ξ0}.

(ii) When the new manager encounters high investor demand, that is, alarger value of z(t), the level of fund-specific alpha δ∗ decreases as afirst-order effect.

(iii) Having a convex learning cost does not affect property (ii).

The extensive margin has a profound effect on the incentive of new managersto generate additional performance. When new managers encounter high in-vestor demand in the search phase (i.e., when z(t) is high), the incentive formanagers to generate additional performance to retain matched investor capi-tal (the intensive margin) declines.12 The interplay between the extensive andintensive margins is one of the most fundamental tradeoffs that new managersface in an economy with search frictions.

This trade-off highlights an important difference between mutual fund andhedge fund startups. In the original BG equilibrium, capital supplied to themutual fund industry is competitive, leading a more skillful fund to enjoy alarger size in realizing its economic rents. In the Internet Appendix, we showthat this positive capital-skill relationship holds even when investors bearsearch costs (and thus compete) for managerial skill—investors still get zeroeconomic rents after search costs. For hedge fund startups, however, managerssearch (and thus compete) for accredited investors. In this case, investors canobtain economic rents, and a negative capital-skill relationship can arise, re-flecting the substitution between the extensive and intensive margins.

We now discuss two concerns related to the generality of Proposition 1. Thefirst is that the new fund may not distribute all fund-specific alpha (δ∗) toinvestors. The retained part can be used to pay for the operation and searchcosts of the fund or can be kept as incentive fees. Lemma IA.1 in the InternetAppendix demonstrates that neither case will change property ii of Proposition1. The intuition is that while both cases change the division of rents betweenthe manager and investors, neither eliminates the interplay between the twomargins during the search-and-bargaining process.

The second concern is how the search-and-bargaining process shapes theincentives of the existing fund. Lemma IA.2 in the Internet Appendix showsthat, under reasonable conditions, the existing fund optimally chooses not todeliver any additional performance. The result is intuitive: as old funds have amuch smaller search burden due to existing investors, they have less incentive

12 The fraction of optimistic investors of the hn type, μhn, can also achieve a similar effect as z(t).However, unlike z(t), μhn is determined by the steady-state of the existing fund and can be furtherinfluenced by an endogenous family structure. Hence, in this proposition we focus on z(t); weexamine μhn in later sections. Thus, in Proposition 1 we assume that new funds are independent(i.e., stand-alone).


to resort to performance as a bargaining tool to attract new capital even whennew funds have the necessity of doing so.

D. Hedge Fund Families and Implications

We now examine the inception of new funds affiliated with the existing fund.The following proposition shows that family structure has a profound effect onthe incentives and performance of new managers in the presence of searchfrictions.

PROPOSITION 2: When an affiliated new manager solves the optimizationproblem in equation (3), the following properties hold:

(i) A family-affiliated inception pays a lower search cost than a stand-aloneinception.

(ii) The optimal level of fund-specific alpha chosen by the family-affiliatedinception is lower than that of a stand-alone inception.

(iii) Property (ii) of Proposition 1 remains valid for family-affiliated incep-tions: high investor demand reduces the level of optimal alpha.

The first property suggests that a positive networking effect reduces thesearch cost of a new fund affiliated with a family, which provides the rationalefor the endogenous emergence of family structure. The next two propertiesdescribe the influence of the family structure on the incentives of a new fundmanager. Since the family structure makes it easier for its affiliated new fundsto find capital in the extensive margin, the incentive to generate performanceto improve the intensive margin is reduced. However, this does not eliminatethe interplay between the extensive and intensive margins. High investor de-mand still reduces performance incentives for an affiliated fund. In the Inter-net Appendix, we further prove that convex learning costs and the retention ofa fraction of rents will not change this demand-performance trade-off.

Note that the proposition above applies to a new fund that has a differentinvestment strategy than its affiliated existing fund—hence investors of thelatter are interested in supplying capital to the new fund as a networking ef-fect. In practice, however, new clone funds are often launched with the samestrategy as the affiliated existing fund(s). The existence of clone funds is puz-zling because a fund family could have asked its existing funds to absorb thecapital instead. What prevents fund families from doing so—and what can wesay about the performance of clone funds?

Lemma IA.3 in the Internet Appendix sheds light on the underlining eco-nomics by examining a second channel through which search frictions giverise to family-affiliated inceptions, namely diseconomies of scale. This channelis widely regarded as a binding constraint for hedge funds (e.g., Goetzmann,Ingersoll, and Ross (2003), Getmansky, Lo, and Makarov (2004), Fung et al.(2008)). In particular, Lemma IA.3 in the Internet Appendix shows that theoptimal search cost paid by an existing fund exhibits a diminishing benefit asfund size increases. The intuition is that funds need to search for new investors


when old investors withdraw. A larger fund size (more search costs spent) re-duces the mass of high-type noninvestors remaining in the market becausethey are converted to high-type investors. This decreases the marginal benefitof spending on search. Since search costs reduce the capital that can other-wise be used to relax the diminishing returns to scale (the BG condition), theyamplify the fund’s diseconomies of scale.

This amplification of diseconomies-of-scale offers one rationale for the incep-tion of clone funds: when an existing fund encounters enthusiastic investors,the launch of a clone fund provides a cost-effective way to retain the capitalfor future periods. Since the goal of the clone fund is to provide an investmentopportunity identical to the existing fund, its performance will not exceed thatof the existing fund. For a clone fund created due to diseconomies of scale con-siderations, its inception reveals the excess demand the fund family faced, re-gardless of whether the clone fund is launched using a hot or cold strategy.Hence, as a first-order effect, we expect clone funds to deliver poorer perfor-mance than nonclones, regardless of whether they arise in hot or cold strategycategories.

E. A Numerical Example and Empirical Hypothesis on Hedge Fund Startups

The predictions of our model about the incentives of new fund managerscan be demonstrated in a numerical example. We set parameters to matchthe observed size and performance of existing and new funds. Specifically,we calibrate the steady-state distribution of investor types (i.e., μσ for σ ∈{ho, lo,hn, ln}) for the existing fund according to Lemma IA.2 in the InternetAppendix, and then apply the corresponding intensity of noninvestors (i.e., μhnand μln) to inceptions according to Proposition 1. Table IA.I in the InternetAppendix tabulates the parameter values, which provide a baseline case to de-scribe the existing and new funds. As can be seen, the baseline case closelymatches the observed size and performance of existing and new funds.

In Figure 1, we show how the incentives of a new hedge fund change when itsextensive margin deviates from the baseline case. Incentives are captured bythe model-implied optimal fund-specific performance that the new fund man-ager is willing to deliver. Variation in the extensive margin is quantified by theratio z(t)/z, where z is the value of z(t) used in the baseline case. We refer tothis ratio as the investor demand index. A high (low) z(t)/z ratio indicates casesin which investors provide more (less) capital to the hedge fund industry. Notethat variation in z(t)/z is exogenous to new managers. This figure plots theoptimal performance, δ∗, of stand-alone inceptions, family-affiliated noncloneinceptions, and family-affiliated clone inceptions according to Proposition 1,Proposition 2, and Lemma IA.3 in the Internet Appendix.

Figure 1 demonstrates the negative relation between investor demand alongthe extensive margin and performance incentives along the intensive mar-gin. This negative relation applies to both stand-alone and family-affiliatedinceptions. Family-affiliated inceptions have less incentive to deliver perfor-mance than stand-alone inceptions. As discussed in the Internet Appendix, the


Figure 1. Optimal alpha of stand-alone and family-affiliated inceptions. This figure plotsthe relationship between investor demand during the search step and inception performance thatthe new fund manager delivers. Variation in investor demand is given by the ratio of z(t)/z, where zis the value of z(t) used in the baseline case of the numerical example provided in the Internet Ap-pendix. The optimal performance of stand-alone inceptions, family-affiliated nonclone inceptions,and family-affiliated clone inceptions is calculated according to Proposition 1, Proposition 2, andLemma IA.3, respectively, in the Internet Appendix.

performance of cold and hot hedge fund inceptions in our empirical analysescan be generated from reasonable ranges of demand changes.

To better describe the empirical implications of the model visualized above,we develop hypotheses that can be tested in the data. According to Proposi-tion 1, our first hypothesis is that the incentive for new managers to deliverperformance differs according to variation in the extensive margin related toinvestor demand. Our model predicts that hot inceptions deliver poorer per-formance than cold inceptions in the cross-section. To test this empirically, wenotice that hot strategies can be proxied by recent high flows and high per-formance (since hedge fund investors chase past performance) in a strategycategory. Accordingly, we have the following hypothesis.

HYPOTHESIS 1 (Two types of inceptions): Hot inceptions (following high cat-egory performance and flows) differ from cold inceptions in their incentive todeliver performance. On a risk-adjusted basis, cold inceptions deliver supe-rior performance in general and outperform hot inceptions in particular.


The null hypothesis is that all hedge fund inceptions are ex ante identicaland, as a result, deliver similar performance ex post. We can also compare theperformance of new funds to that of existing funds. The two propositions aboveand Lemma IA.2 in the Internet Appendix imply that only managers of coldinceptions have the incentive to deliver performance above and beyond thebenchmark performance of old funds. This novel heterogeneity is summarizedin the following hypothesis.

HYPOTHESIS 2 (Value-creating inceptions): Inceptions are value-creatingand associated with better performance than existing funds. However, thisvalue-creation effect concentrates in cold inceptions only.

Our model suggests that family structure arises endogenously in a marketwith search frictions, and influences the performance of inceptions. Proposition2 predicts that family structure affects the performance incentives of affiliatednonclone inceptions compared to stand-alone inceptions. Lemma IA.3 in the In-ternet Appendix suggests that search-friction-amplified diseconomies of scalegive rise to the inception of clone funds. These predictions can be summarizedin the following hypotheses.

HYPOTHESIS 3 (The impact of family structure on inception performance):Hedge fund inceptions within existing families deliver poorer performancethan stand-alone inceptions.

HYPOTHESIS 4 (Two types of inceptions within family-affiliated funds):Within family-affiliated nonclone funds, cold inceptions outperform hot in-ceptions. Clone inceptions, by contrast, deliver poor performance regardlessof being launched in cold or hot categories.

The two hypotheses above propose that search-friction-motivated familystructures critically influence inception performance. In the mutual fund lit-erature, researchers have identified several important mechanisms for fam-ily structure, including the convenience of reduced within-family switchingfees (Massa (2003)), the efficiency of within-family resource allocation (Fang,Kempf, and Trapp (2014), Berk, van Binsbergen, and Liu (2017)), the flexi-bility of cross-subsidization (e.g., Bhattacharya, Lee, and Pool (2013), amongothers), and the star-creation strategy of attracting flows (Nanda, Wang, andZheng (2004)). Our model suggests that the hedge fund industry is insteaddominated by search frictions and associated with a different rationale for thecreation of fund families.

II. Data Description

We use monthly hedge fund data formed by merging fund and return in-formation from three sources: Lipper TASS, HFR, and BarclayHedge. Fundcounts for our merged database, the constituent databases, and the degree ofoverlap are reported in Table IA.II in the Internet Appendix. After the merge,we have a total of 31,402 funds. Our sample includes both live and defunctfunds to mitigate survivorship bias and spans the 1994 to 2016 period. We


restrict attention to funds reporting at least 12 monthly return observationsand therefore retrieve inception information from the merged database up tothe end of 2015. Funds reporting the same management company are consid-ered to be from the same fund family. Because funds may come from differentdatabases, we cross-reference family affiliations across databases.

Because funds have the option of reporting backfilled returns at the timethey start reporting to a commercial database, hedge fund data are prone tobackfill bias. We use the method proposed in Jorion and Schwarz (2019) to esti-mate the date when each fund began reporting returns based on the sequentialassignment of fund IDs within databases. This is necessary for BarclayHedge,which does not provide information on which returns are backfilled. In addi-tion, TASS has not updated the fund add-date field since 2011, so we apply thismethod to TASS returns where necessary. HFR reports complete and reliableinformation on fund add dates, so estimation is not necessary. We minimize theimpact of backfill bias by excluding fund returns prior to the fund’s add date.Each database also reports each fund’s inception date.

Unless otherwise specified, we report returns in excess of the risk-free rate.We use the seven-factor model from Fung and Hsieh (2001, 2004) to computethe risk-adjusted return (alpha). The factors are constructed following the in-structions from David Hsieh’s Hedge Fund Data Library.13

The databases provide strategy classifications for each fund, but eachdatabase uses a different categorization methodology. For the purposes of thispaper, we use the TASS strategy classification criteria. TASS provides 10 gen-eral hedge fund strategies—CA, dedicated short bias (DS), event-driven (ED),emerging markets (EM), equity market neutral (EMN), fixed income arbitrage(FI), global macro (GM), long/short equity (LS), managed futures (MF), andmulti-strategy (MS)—which provide a reasonable cross-section to examine in-ception incentives.14 To map reported strategy categories from HFR and Bar-clayHedge to the TASS definitions, we use fund merge information. For eachfund that appears in both TASS and one of the other databases, we record themapping from the database classification into the TASS classification. Eachdatabase classification can then be assigned to a TASS strategy based on a ma-jority relationship. The resulting strategy mapping is reported in Table IA.IIIin the Internet Appendix.

Table I reports the number of funds reporting a valid AUM at the end of eachyear in selected hedge fund strategies. Section III in the Internet Appendixprovides a more complete version of the table, which reports the number foreach of the 10 hedge fund strategies. We exclude funds reporting other, minor,strategies and funds of funds. The total number of funds has steadily increasedover time. We also report the number of distinct families. Over time, the aver-age number of funds per family has increased from 1.67 at the beginning of our

13 See https://faculty.fuqua.duke.edu/∼dah7/HFRFData.htm.14 The two other databases have, relatively speaking, too broad tier-one strategy categories

and too detailed tier-two strategies for our testing purposes. In contrast, the TASS classificationachieves a sensible balance between the number of strategies and the number of funds within eachstrategy.

https://faculty.fuqua.duke.edu/%7Edah7/HFRFData.htm

16 The Journal of Finance®T

able

IS

um

mar

yS

tati

stic

sof

Hed

geF

un

ds

and

Hed

geF

un

dF

amil

ies

Th

ista

ble

repo

rts

the

nu

mbe

rof

fun

dsin

our

sam

ple

byin

vest

men

tst

rate

gyca

tego

ryan

dye

ar.F

un

dsar

ein

clu

ded

ifth

eyre

port

are

turn

inD

ecem

ber

ofth

egi

ven

year

and

hav

eat

leas

t12

mon

thly

retu

rnob

serv

atio

ns.

Str

ateg

ycl

assi

fica

tion

sar

eco

nve

rtib

lear

bitr

age

(CA

),de

dica

ted

shor

tbi

as(D

S),

even

t-dr

iven

(ED

),em

ergi

ng

mar

kets

(EM

),eq

uit

ym

arke

tn

eutr

al(E

MN

),fi

xed

inco

me

arbi

trag

e(F

I),

glob

alm

acro

(GM

),lo

ng/

shor

teq

uit

y(L

S),

man

aged

futu

res

(MF

),an

dm

ult

i-st

rate

gy(M

S).

We

repo

rtth

en

um

ber

offu

nds

for

CA

,E

MN

,G

M,

and

LS,

and

refe

rth

ere

ader

toS

ecti

onII

Iof

the

Inte

rnet

App

endi

xfo

ra

vers

ion

ofth

eta

ble

incl

udi

ng

all

stra

tegi

es.W

eex

clu

defu

nds

offu

nds

and

fun

dsw

ith

out

stra

tegy

info

rmat

ion

.We

also

repo

rtth

eto

tal

nu

mbe

rof

hed

gefu

nds

and

hed

gefu

nd

fam

ilie

s,as

wel

las

the

aver

age

nu

mbe

rof

fun

dspe

rfa

mil

yan

dth

efr

acti

onof

fam

ilie

sth

ath

ave

mu

ltip

lefu

nds

atth

een

dof

each

year

.Ou

rsa

mpl

eis

form

edby

mer

gin

gfu

nd/

fam

ily/

retu

rnin

form

atio

nfr

omT

AS

S,H

FR

,an

dB

arcl

ayH

edge

and

span

sth

epe

riod

1994

to20

16.

Str

ateg

yC

ateg

ory

Cou

nts

Year

CA

EM

NG

ML

SN

um

ber

ofF

un

dsN

um

ber

ofFa

mil

ies

Fu

nds

per

Fam

ily

Fra

ctio

nM

ult

i-F

un

d

1994

7772

297

781

2,59

21,

555

1.67

0.30

1995

9189

338

960

2,99

61,

722

1.74

0.33

1996

106

9832

01,

195

3,33

61,

890

1.77

0.34

1997

114

136

328

1,39

13,

724

2,05

21.

810.

3619

9813

517

933

91,

599

4,03

22,

215

1.82

0.36

1999

153

217

330

1,89

24,

465

2,37

41.

880.

3720

0019

023

834

02,

154

4,88

52,

497

1.96

0.39

2001

239

321

375

2,47

65,

527

2,71

12.

040.

4020

0228

940

945

62,

706

6,25

92,

883

2.17

0.41

2003

320

459

609

3,13

17,

345

3,21

02.

290.

4320

0431

251

373

43,

616

8,53

53,

512

2.43

0.44

2005

294

578

840

4,12

89,

655

3,83

52.

520.

4420

0628

059

593

64,

558

10,6

124,

042

2.63

0.44

2007

239

601

991

4,74

211

,196

4,16

72.

690.

4420

0817

854

41,

015

4,45

110

,765

4,07

02.

640.

4320

0915

653

21,

140

4,49

411

,099

4,10

92.

700.

4220

1018

049

41,

194

4,60

211

,611

4,12

42.

820.

4320

1119

049

31,

208

4,68

311

,800

4,01

52.

940.

4420

1220

046

41,

186

4,63

111

,554

3,87

72.

980.

4320

1320

643

91,

016

4,65

110

,966

3,74

32.

930.

4220

1419

646

292

44,

485

10,4

103,

526

2.95

0.42

2015

172

439

792

4,14

19,

418

3,20

52.

940.

4120

1615

937

965

33,

583

8,06

52,

812

2.87

0.40


sample to 2.87 by the end. The fraction of management families with multiplefunds has also increased, from 30% in 1994 to 40% at the end of 2016. Both ob-servations suggest that family structure plays an increasingly important rolein the hedge fund industry.

Table II reports total inceptions per year and total funds reporting in De-cember of the given year. The proportion of the universe represented by newfunds increased from 18% in 1994 to 22% in 2003 and then decreased there-after. We also report the total AUM of new funds raised each year (InceptionAUM) and the AUM of our whole sample. We consider the inception AUM tobe the first nonmissing reported AUM in the first three months of its life. In-ception AUM grew from $2.8 billion in 1994 to $47.1 billion in 2006 and wasvolatile thereafter. We also report flows to existing funds each year. Flows arecomputed from performance and AUM according to

Flowi,t = AUMi,t − AUMi,t−1 · (1 + ri,t ), (4)

where fund flows and AUM are reported in U.S. dollars (we convert any AUMreported in another currency to U.S. dollars). The variable ri,t represents thereturn to fund i in month t. Compared to flows to existing funds, inceptionAUM is much less volatile, suggesting that new hedge funds play a unique andimportant role in attracting capital to the hedge fund industry.

In the sixth column, we report the number of inceptions that are stand-alone.The difference between the first column and this column indicates the numberof inceptions affiliated with an existing family. In the last column, among thefamily-affiliated inceptions, we report how many are clones of other funds. Aninception in an existing family is categorized as a clone if it is in the samestrategy category as an existing fund in the family and if the fund has a returncorrelation with the previously existing fund of 90% or greater. Where treatedseparately, a “nonclone” inception in an existing family is the first fund in astrategy within that family. Overall, 10,620 inceptions in our sample were thefirst in their management companies and 17,564 are inceptions in existingfamilies. Of the inceptions in existing families, 8,950 are classified as clonefunds. Overall, our sample contains 28,184 inceptions.

Using monthly returns, we construct the raw and risk-adjusted performanceof funds and portfolios of funds. In each case, we measure raw performance bycomputing the 60-month excess returns of each fund over the risk-free rate.Risk-adjusted returns are computed as the intercept (alpha) from a 60-monthregression of fund excess returns on the seven hedge fund risk factors proposedby Fung and Hsieh (2004). The regression equation is

rp,t = αp + βp,1MKTt + βp,2SMBt + βp,3YLDCHGt + βp,4BAAMTSYt+ βp,5PTFSBDt + βp,6PTFSFXt + βp,7PTFSCOMt + εp,t,

(5)

where rp,t is the monthly excess return to portfolio p in month t, MKT is the ex-cess return to the market, SMB is the small-minus-big size factor, YLDCHG isthe change in the 10-year Treasury constant-maturity yield, BAAMTSY is thechange in Moody’s Baa yield less the 10-year Treasury constant-maturity yield,


Tab

leII

Su

mm

ary

Sta

tist

ics

ofH

edge

Fu

nd

Ince

pti

ons

and

Flo

ws

over

Tim

eW

ere

port

flow

sto

hed

gefu

nd

ince

ptio

ns

and

exis

tin

gfu

nds

byye

ar.

Var

iabl

esin

clu

deth

en

um

ber

ince

ptio

ns,

and

the

rati

oof

ince

ptio

ns

toal

lex

isti

ng

fun

dsin

Dec

embe

r,in

cept

ion

AU

M(c

ompu

ted

asth

efi

rst

AU

Min

the

firs

tth

ree

mon

ths

for

each

fun

d),

and

net

flow

sto

all

fun

dsin

this

year

(com

pute

d,fu

nd

byfu

nd,

asF

low

=A

UM

t−

AU

Mt−

1(1

+r t

)).T

otal

AU

Mis

the

AU

Mof

all

fun

dsin

Dec

embe

r.S

tan

d-al

one

ince

ptio

nis

the

nu

mbe

rof

ince

ptio

ns

that

begi

nn

ewfa

mil

ies,

and

clon

ein

cept

ion

sar

ein

cept

ion

sin

exis

tin

gfa

mil

ies

inth

esa

me

stra

tegy

and

90%

corr

elat

edw

ith

apr

evio

usl

yex

isti

ng

fun

din

the

fam

ily.

Ou

rsa

mpl

esp

ans

from

1994

to20

16bu

tfu

nds

are

requ

ired

toh

ave

12m

onth

lyre

turn

obse

rvat

ion

sfo

rin

clu

sion

,so

our

ince

ptio

nsa

mpl

een

dsin

Dec

embe

r20

15.

Year

All

Ince

ptio

ns

Ince

ptio

ns/

Exi

stin

gIn

cept

ion

AU

M(B

il)

Net

Flo

ws

(Bil

)T

otal

AU

M(B

il)

Sta

nd-

Alo

ne

Ince

ptio

ns

Clo

ne

Ince

ptio

ns

1994

479

0.18

2.82

9.19

103.

9829

651

1995

575

0.19

4.36

–2.8

912

9.31

282

117

1996

653

0.20

6.65

7.01

169.

4637

410

419

9768

40.

186.

9932

.08

247.

3835

413

519

9869

90.

178.

6217

.22

277.

1339

012

819

9988

90.

2013

.35

7.06

344.

7046

419

720

0089

10.

1810

.94

9.79

381.

7043

818

120

011,

086

0.20

11.1

960

.50

451.

4951

027

820

021,

320

0.21

14.9

939

.99

480.

2659

633

720

031,

597

0.22

31.2

613

8.85

731.

4164

747

320

041,

815

0.21

34.0

023

3.53

1,03

4.09

675

613

2005

1,96

50.

2041

.92

35.6

71,

164.

7773

362

920

062,

013

0.19

47.1

123

4.18

1,48

4.61

715

667

2007

1,96

60.

1842

.90

255.

431,

865.

7466

367

120

081,

613

0.15

39.1

2–2

07.8

81,

259.

3462

541

820

091,

839

0.17

58.8

9–8

8.19

1,32

1.85

570

644

2010

1,82

50.

1669

.44

61.0

41,

522.

6653

571

820

111,

711

0.14

71.6

540

.58

1,60

3.76

476

694

2012

1,52

70.

1397

.04

–4.5

81,

678.

6945

660

020

131,

309

0.12

53.0

911

9.88

1,87

7.84

370

528

2014

1,05

50.

1041

.19

10.6

22,

003.

0928

546

020

1567

30.

0726

.70

71.7

22,

216.

7316

630

7


and the other three variables are trend-following factors available on Hsieh’swebsite: PTFSBD (bond), PTFSFX (currency), and PTFSCOM (commodity).

III. Determinants of Hedge Fund Inception Probability

Before categorizing inceptions as cold or hot, we examine how various char-acteristics of hedge fund strategy classifications and fund families affect theincentives of inceptions. This analysis provides intuition to help us identifydifferent types of inceptions.

We start with a logistic regression specification of the incidence of hedgefund inception by date, strategy category, and family, linking the incentives oflaunching new hedge funds to a list of category and family characteristics. Thedependent variable is set to 1 when a family had an inception in a given yearand strategy category and 0 otherwise. The regression equation is

Inceptionj,k,t = (α + β × Xj,t−1 + ψ × Yk,t−1

) + ε j,k,t, (6)

where (·) represents the logistic function, Xj,t−1 is a vector of strategy ex-planatory variables for strategy category j and year t – 1, Yk,t−1 a vector offamily explanatory variables for family k and year t – 1, Strategy return is theaverage monthly return of an equal-weighted portfolio of funds in each strat-egy category in year t – 1, Strategy volatility is computed from equal-weightedportfolios of funds over the 24 months prior to year t, Strategy AUM is the sumof reported AUM in December of year t – 1, Strategy inceptions is the numberof inceptions by strategy category j in year t – 1, normalized by the number offunds in strategy category j at the end of year t – 2, and Family return, Familyvolatility, and Family AUM are defined following the corresponding strategyvariables. In addition, Family assets in same strategy is the sum of assets instrategy j and family k at the end of year t – 1, Strategy large family open is setto 1 if one of the largest eight hedge fund families had an inception in strategycategory j in year t – 1, Family inceptions is the count of inceptions in familyk in year t – 1. In all models, year fixed effects are included as yearly dummyvariables in the regression.

Table III reports the estimation results. Specifications (1) through (3) focuson strategy-level controls. Specifications (4) through (6) add family-level con-trols. In specification (7), we include an interaction term between the returnsof the strategy and the assets that a family already has in that strategy. Thistest examines whether there is a nonlinear effect of the family already havinga strong presence in a given strategy at the time that the strategy becomespopular. This would be a likely time for a family to initiate a fund to takeadvantage of investor demand for that strategy.

The estimation results in specifications (1) and (2) show that lagged strat-egy returns and lagged strategy flows are positively associated with inceptionsin a family/year/strategy, with coefficients of 24.2 and 11.8, respectively (t-statistic = 14.15 and 7.44). Lagged inceptions in the strategy are also related toinceptions (t-statistic = 19.09) in specification (3). These results are highly


Tab

leII

IL

ogis

tic

Reg

ress

ion

ofIn

cep

tion

sw

ith

inE

xist

ing

Fam

ilie

son

Fam

ily/

Str

ateg

yV

aria

ble

sT

his

tabl

ere

port

slo

gist

icre

gres

sion

resu

lts

ofth

edu

mm

yva

riab

lefo

rw

het

her

ther

ew

asan

ince

ptio

nin

agi

ven

fam

ily/

year

/str

ateg

yon

char

ac-

teri

stic

sof

that

fam

ily,

year

,an

dst

rate

gyca

tego

ry.T

he

logi

stic

regr

essi

oneq

uat

ion

isIn

cept

ion

j,k,

t=

(α+β

×X

j,t−1

+ψ

×Y

k,t−

1)+

εj,

k,t,

wh

ere

(·)

repr

esen

tsth

elo

gist

icli

nk

fun

ctio

n,X

j,t−1

isa

vect

orof

stra

tegy

-spe

cifi

cva

riab

les

for

stra

tegy

cate

gory

jin

year

t–

1an

dY

k,t-

1is

ave

ctor

offa

mil

y-sp

ecifi

cva

riab

les

for

fam

ily

kin

year

t–

1;In

cept

ion

j,k,

tis

adu

mm

yva

riab

leth

atis

1if

ther

ew

asan

ince

ptio

nin

stra

tegy

j,fa

mil

yk,

and

year

t,an

d0

oth

erw

ise.

Inye

art,

the

expl

anat

ory

vari

able

sar

eas

foll

ows.

Str

ateg

yre

turn

and

Fam

ily

retu

rnar

eth

eav

erag

em

onth

lyre

turn

toan

equ

al-w

eigh

ted

port

foli

oof

fun

dsfo

ra

give

nst

rate

gyor

fam

ily

inye

art

–1.

Str

ateg

yvo

lati

lity

and

Fam

ily

vola

tili

tyar

eco

mpu

ted

from

mon

thly

equ

al-w

eigh

ted

port

foli

osfr

omt

–2

tot.

Str

ateg

yA

UM

and

Fam

ily

AU

Mar

eth

esu

mof

AU

Min

Dec

embe

rof

year

t–

1in

bill

ion

sof

US

D.F

amil

yas

sets

insa

me

stra

tegy

isth

esu

mof

asse

tsin

stra

tegy

jan

dfa

mil

yk

atth

een

dof

year

t–

1.S

trat

egy

ince

ptio

ns

isth

en

um

ber

ofin

cept

ion

sin

stra

tegy

jin

year

t–

1,n

orm

aliz

edby

the

nu

mbe

rof

fun

dsin

stra

tegy

jat

the

end

ofye

art

–2.

Str

ateg

yla

rge

fam

ily

open

isse

tto

1if

one

ofth

ela

rges

tei

ght

hed

gefu

nd

fam

ilie

sh

adan

ince

ptio

nin

stra

tegy

jin

year

t–

1.F

amil

yin

cept

ion

sis

the

cou

nt

ofin

cept

ion

sin

fam

ily

kin

year

t–

1.In

allm

odel

s,ye

arfi

xed

effe

cts

are

incl

ude

das

year

lydu

mm

yva

riab

les

inth

ere

gres

sion

.t-S

tati

stic

sar

ein

pare

nth

eses

.Sta

tist

ical

sign

ifica

nce

atth

e1%

,5%

,an

d10

%le

vels

isde

not

edby

***,

**,a

nd

*,re

spec

tive

ly.

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Str

ateg

yre

turn

24.1

77**

*24

.403

***

22.6

79**

*(1

4.15

)(1

3.63

)(1

2.66

)S

trat

egy

flow

11.7

70**

*11

.965

***

10.5

87**

*(7

.44)

(7.5

1)(6

.54)

Str

ateg

yin

cept

ion

s2.

179*

**2.

238*

**(1

9.09

)(1

9.29

)S

trat

egy

vola

tili

ty2.

507*

**3.

436*

**3.

615*

**2.

917*

**3.

867*

**4.

104*

**3.

346*

**(6

.98)

(9.9

7)(1

0.34

)(7

.95)

(11.

00)

(11.

50)

(9.1

4)S

trat

egy

AU

M(B

n)

0.00

0***

0.00

0***

0.00

0***

0.00

0***

0.00

0***

0.00

0***

0.00

0***

(25.

48)

(24.

26)

(27.

53)

(24.

16)

(23.

33)

(26.

55)

(24.

69)

Str

ateg

yla

rge

fam

ily

open

0.53

8***

0.57

5***

0.49

9***

0.54

1***

0.57

5***

0.49

7***

0.53

4***

(14.

14)

(15.

11)

(13.

01)

(14.

02)

(14.

89)

(12.

78)

(13.

79)

Fam

ily

retu

rn0.

058

3.28

6***

2.33

2**

−0.0

83(0

.06)

(3.5

4)(2

.49)

(−0.

09)

Fam

ily

ince

ptio

ns

0.13

0***

0.13

0***

0.13

1***

0.12

9***

(36.

01)

(36.

01)

(36.

12)

(35.

89)

(Con

tin

ued

)


Tab

leII

I—C

onti

nu

ed

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Fam

ily

vola

tili

ty−0

.346

**−0

.387

***

−0.4

98**

*−0

.310

**(−

2.31

)(−

2.58

)(−

3.30

)(−

2.07

)F

amil

yA

UM

(Bn

)−0

.007

**−0

.007

**−0

.006

**−0

.004

(−2.

29)

(−2.

29)

(−2.

03)

(−1.

45)

Fam

ily

flow

0.34

6***

0.33

7***

0.33

1***

0.33

4***

(11.

90)

(11.

55)

(11.

36)

(11.

41)

Fam

ily

asse

tsin

sam

est

rate

gy(B

n)

0.06

8***

0.06

7***

0.06

6***

0.03

3***

(8.2

3)(8

.21)

(8.0

7)(3

.64)

Str

ateg

yre

turn

×fa

mil

yas

sets

9.61

2***

(6.3

0)

Pse

udo

-R2

2.70

%2.

40%

2.80

%6.

00%

5.60

%6.

10%

6.10

%


robust in specifications (4) through (7). Other strategy-level variables thatare positively associated with subsequent inceptions include volatility, strat-egy AUM, and the strategy having a recent inception from a very large family.

The positive relations between strategy return, flow, and inceptions and sub-sequent inceptions are relevant to our analysis. We employ lagged strategy re-turns and lagged strategy flow as proxies for investor demand for that strategy.We use these proxies to identify cold and hot strategies and use lagged strategyinceptions as an alternative proxy in a robustness check.

Although family flows are related to the inception of new funds (see specifica-tion (7)), family returns are insignificant after we control for strategy returns.This difference between family and strategy returns suggests that the familystructure and strategy-level demand operate on inceptions through differentchannels. The positive relation between family assets in a strategy and subse-quent inceptions in the strategy (i.e., clone inceptions) reveals that clone fundscan be launched when families are unwilling or unable to use existing funds toefficiently absorb more capital due to diseconomies of scale. Finally, specifica-tion (7) reports an interaction effect: families with high assets in the previousyear in strategy categories that had good returns are particularly likely to havean inception (t-statistic = 6.30).

Taken together, both strategy-level variables related to investor demand andthe presence of a family structure facilitate the inception of new hedge funds.These results motivate our identification strategies, which we use to examinethe performance of the two types of inceptions.

IV. Performance Difference between Cold and Hot Inceptions

In this section, we test our hypotheses by constructing cold and hot inceptionportfolios and examining their performance differences.

A. Inceptions in Hot and Cold Strategy Categories

We start by analyzing inception performance in hot and cold strategy cate-gories. To determine whether an inception is in a hot or cold strategy, we usetwo measures: the 36-month (prior to inception) flows into a given strategyand the 36-month (prior to inception) returns to the strategy. Each month, werank the 10 hedge fund strategies using these two lagged variables. Strategycategories with a high rank (eight or greater) in both measures are defined as“hot,” whereas strategies with a low rank (three or lower) in both measuresare defined as “cold.” New hedge funds employing a hot (cold) strategy areaccordingly classified as hot (cold) inceptions.

In each month, portfolios are formed from new hedge fund inceptions overthe prior three-month period. Among these inceptions, cold and hot inceptionsare identified and included in their respective inception portfolios while ex-isting funds alive during that period are included in the noninception portfo-lio. Each inception is held in its portfolio for a 60-month holding period afterits inception. Holding periods follow actual fund inception dates. We exclude


backfilled returns and require at least 12 monthly return observations for aninception to be included. Funds within portfolios are equally weighted. Sincethe inception portfolios of a given type are created in each month and heldfor 60 months to assess their performance, we follow Jegadeesh and Titman(1993) and equal-weight overlapped inception portfolios in each month to cre-ate the final holding portfolio for the inception type. Portfolio returns are thenregressed on the seven Fung and Hsieh (2004) risk factors to obtain the risk-adjusted alpha (see equation (5) for the regression specification).

These portfolio regression results are in the first three columns of Table IV.The portfolio of cold inceptions has a significant alpha of 0.376% monthly or4.6% per year (t-statistic = 4.95), as can be seen in the first column, whereasthat of hot inceptions in the second column does not have a significant alpha(t-statistic = 1.1). Furthermore, the monthly spread between the two portfo-lios (labeled “Cold-Hot”) is 0.242% per month, or 2.9% per year (t-statistic =2.36). These estimates are economically sizable and support our first hypothe-sis, suggesting that cold inceptions deliver superior risk-adjusted performancein general and outperform hot inceptions in particular.

The next three columns report the performance of existing funds (labeled“Old funds”), as well as the spread between cold/hot inceptions and existingfunds. Existing funds also deliver significant risk-adjusted performance (α =0.224% per month, or 2.7% annually). This result is consistent with the litera-ture documenting that hedge funds deliver abnormal performance (see, amongothers, Fung and Hsieh (1997), Ackermann, McEnally, and Ravenscraft (1999),Agarwal and Naik (2004), Getmansky, Lo, and Makarov (2004), Kosowski,Naik, and Teo (2007), Agarwal, Daniel, and Naik (2009, 2011), Aragon andNanda (2012), Sun, Wang, and Zheng (2012), Cao et al. (2013), Jiao, Massa,and Zhang (2016)).

More importantly for our purposes, the last two columns show that cold in-ceptions outperform existing funds by 0.152% per month or 1.8% per year (t-statistic = 2.35), whereas hot inceptions do not deliver a significant alpha overexisting funds (nor do they deliver a significant alpha overall). Hence, our re-sults support Hypothesis 2, which predicts that the value-creation effect of newfunds is concentrated in cold inceptions.

B. Family-Affiliated Inceptions in Hot and Cold Strategy Categories

We now examine the effect of family structure on inception performance.We first examine the general difference between stand-alone inceptions (ornew family inceptions, as each inception effectively creates a new family) andfamily-affiliated inceptions. For each group of inceptions, we form inceptionportfolios as above and regress portfolio returns (and pairwise and cornerspreads) on hedge fund risk factors to obtain the risk-adjusted alpha.

The alphas of stand-alone inceptions and family-affiliated inceptions are re-ported in Table V. Stand-alone inceptions generate a risk-adjusted alpha of0.458% per month (5.7% annually) with a t-statistic of 8.49. The alpha offamily-affiliated inceptions is smaller, at 0.233% per month (2.8% annually),

24 The Journal of Finance®T

able

IVP

erfo

rman

ceD

iffe

ren

ces

bet

wee

nP

ortf

olio

sb

yIn

cep

tion

Typ

eT

his

tabl

epr

esen

tsth

ere

sult

sof

regr

essi

onan

alys

isof

ince

ptio

ns

and

exis

tin

gfu

nds

.W

eco

mpa

reth

epe

rfor

man

ceof

ince

ptio

ns

inh

ot(p

opu

lar)

stra

tegi

esw

ith

thos

ein

cold

(un

popu

lar)

stra

tegi

esas

wel

las

the

lon

g-sh

ort

spre

adpo

rtfo

lio.

We

also

exam

ine

the

port

foli

oof

alli

nce

ptio

ns

and

the

old-

fun

dpo

rtfo

lio.

Ast

rate

gyis

clas

sifi

edas

hot

(col

d)if

its

past

36-m

onth

retu

rns

and

flow

sar

ebo

tham

ong

the

top

(bot

tom

)30%

ofal

lstr

ateg

ies.

Inan

ym

onth

,in

cept

ion

port

foli

osar

efo

rmed

from

new

hed

gefu

nd

ince

ptio

ns

ofa

give

nfa

mil

yst

ruct

ure

and

stra

tegy

iden

tifi

cati

onov

erth

epr

ior

thre

em

onth

s.E

ach

ince

ptio

nw

ill

beh

eld

init

sco

rres

pon

din

gpo

rtfo

lio

for

a60

-mon

thh

oldi

ng

peri

odaf

ter

its

ince

ptio

n.

Th

eh

oldi

ng

peri

odfo

llow

sth

eac

tual

ince

ptio

nda

teof

each

fun

d.W

ith

inth

eh

oldi

ng

peri

od,w

eex

clu

deba

ckfi

lled

retu

rns

and

requ

ire

atle

ast

12m

onth

lyre

turn

obse

rvat

ion

sfo

ran

ince

ptio

nto

bein

clu

ded

inan

yin

cept

ion

port

foli

o.F

un

dsar

eeq

ual

lyw

eigh

ted

and

reba

lan

ced

atth

ebe

gin

nin

gof

each

mon

th.P

ortf

olio

retu

rns

are

then

regr

esse

don

the

Fu

ng

and

Hsi

eh(2

004)

seve

nri

skfa

ctor

sto

obta

inth

eri

sk-a

dju

sted

retu

rn(a

lph

a).T

he

regr

essi

oneq

uat

ion

is

r p,t

=α

p+β

p,1M

KT t

+β

p,2S

MB

t+β

p,3Y

LD

CH

Gt+β

p,4B

AA

MT

SY

t+β

p,5P

TF

SB

Dt+β

p,6P

TF

SF

Xt+β

p,7P

TF

SC

OM

t+ε

p,t,

wh

ere

r p,t

isth

eex

cess

retu

rnon

port

foli

op

inm

onth

t.T

he

inde

pen

den

tva

riab

les

are

the

mar

ket

exce

ssre

turn

(MK

T),

asi

zefa

ctor

(SM

B),

the

mon

thly

chan

gein

the

10-y

ear

Tre

asu

ryco

nst

ant-

mat

uri

tyyi

eld

(YL

DC

HG

),th

em

onth

lych

ange

inM

oody

’sB

aayi

eld

less

the

10-y

ear

Tre

a-su

ryco

nst

ant-

mat

uri

tyyi

eld

(BA

AM

TS

Y),

and

thre

etr

end-

foll

owin

gfa

ctor

s:P

FT

SB

D(b

ond)

,P

FT

SF

X(c

urr

ency

),an

dP

FT

SC

OM

(com

mod

ity)

.t-

Sta

tist

ics

are

inpa

ren

thes

es.S

tati

stic

alsi

gnifi

can

ceat

the

1%,5

%,a

nd

10%

leve

lsis

den

oted

by**

*,**

,an

d*,

resp

ecti

vely

.

Col

dIn

cept

ion

Por

tfol

io

Hot

Ince

ptio

nP

ortf

olio

Col

d–

Hot

Ince

ptio

nP

ortf

olio

All

Ince

ptio

nP

ortf

olio

Old

Fu

nd

Por

tfol

io

Col

dIn

cept

ion

–O

ldF

un

dP

ortf

olio

Hot

Ince

ptio

n–

Old

Fu

nd

Por

tfol

io

alph

a0.

376*

**0.

134

0.24

2**

0.31

3***

0.22

4***

0.15

2**

−0.0

90(4

.95)

(1.1

1)(2

.36)

(5.4

6)(3

.53)

(2.3

5)(−

0.90

)M

KT

0.20

2***

0.32

0***

−0.1

18**

*0.

261*

**0.

269*

**−0

.066

***

0.05

2**

(10.

79)

(10.

75)

(−4.

64)

(18.

43)

(17.

16)

(−4.

13)

(2.0

8)S

MB

0.10

2***

0.05

10.

051

0.13

4***

0.13

4***

−0.0

32−0

.083

***

(4.3

9)(1

.39)

(1.6

2)(7

.61)

(6.8

9)(−

1.59

)(−

2.68

)Y

LD

CH

G−0

.876

**−0

.039

−0.8

37−0

.487

*−0

.585

*−0

.291

0.54

6(−

2.27

)(−

0.06

)(−

1.61

)(−

1.68

)(−

1.82

)(−

0.88

)(1

.07)

BA

AM

TS

Y−3

.137

***

−3.1

49**

*0.

012

−2.2

27**

*−2

.672

***

−0.4

65−0

.477

(−6.

35)

(−4.

02)

(0.0

2)(−

5.99

)(−

6.49

)(−

1.11

)(−

0.73

)P

TF

SB

D0.

005

−0.0

14*

0.01

9***

−0.0

010.

000

0.00

5−0

.014

**(0

.92)

(−1.

69)

(2.6

6)(−

0.28

)(0

.05)

(1.0

4)(−

2.05

)P

TF

SF

X0.

007

0.02

5***

−0.0

18**

*0.

010*

**0.

016*

**−0

.009

**0.

010*

(1.6

0)(3

.68)

(−3.

13)

(2.9

3)(4

.30)

(−2.

32)

(1.7

0)P

TF

SC

OM

−0.0

01−0

.010

0.00

90.

004

0.00

8*−0

.009

*−0

.019

**(−

0.12

)(−

1.14

)(1

.24)

(0.9

3)(1

.82)

(−1.

92)

(−2.

51)

Adj

ust

edR

246

.20%

41.1

0%12

.70%

67.3

0%64

.50%

8.10

%7.

70%


Tab

leV

Th

eIn

flu

ence

ofF

amil

yS

tru

ctu

reon

Ince

pti

onP

erfo

rman

ceT

his

tabl

epr

esen

tsre

sult

sof

the

anal

ysis

ofth

ein

flu

ence

offa

mil

yst

ruct

ure

onin

cept

ion

s.W

efo

rmin

cept

ion

port

foli

osba

sed

on(1

)th

efa

mil

yst

ruct

ure

ofea

chin

cept

ion

(i.e

.,th

est

and-

alon

ein

cept

ion

,or

fam

ily-

affi

liat

edin

cept

ion

incl

udi

ng

non

clon

ein

cept

ion

san

dcl

one

ince

ptio

ns)

,an

d(2

)th

est

rate

gy-b

ased

iden

tifi

cati

onof

each

ince

ptio

n(i

.e.,

cold

orh

otin

cept

ion

).In

any

mon

th,

ince

ptio

npo

rtfo

lios

are

form

edfr

omn

ewh

edge

fun

din

cept

ion

sof

agi

ven

fam

ily

stru

ctu

rean

dst

rate

gyid

enti

fica

tion

over

the

prio

rth

ree

mon

ths.

Eac

hin

cept

ion

ish

eld

init

sco

rres

pon

din

gpo

rtfo

lio

for

a60

-mon

thh

oldi

ng

peri

odaf

ter

its

ince

ptio

n.T

he

hol

din

gpe

riod

foll

ows

the

actu

alin

cept

ion

date

ofea

chfu

nd.

Wit

hin

the

hol

din

gpe

riod

,we

excl

ude

back

fill

edre

turn

san

dre

quir

eat

leas

t12

mon

thly

retu

rnob

serv

atio

ns

for

anin

cept

ion

tobe

incl

ude

din

any

ince

ptio

npo

rtfo

lio.

Fu

nds

are

equ

ally

wei

ghte

dan

dre

bala

nce

dat

the

begi

nn

ing

ofea

chm

onth

.Por

tfol

iore

turn

sar

eth

enre

gres

sed

onth

eF

un

gan

dH

sieh

(200

4)se

ven

fact

ors

toob

tain

the

risk

-adj

ust

edal

pha.

Th

ere

gres

sion

equ

atio

nis

pres

ente

din

Tab

leIV

.t-S

tati

stic

sar

ein

pare

nth

eses

.Sta

tist

ical

sign

ifica

nce

atth

e1%

,5%

,an

d10

%le

vels

isde

not

edby

***,

**,a

nd

*,re

spec

tive

ly.

Sta

nd-

Alo

ne

(New

Fam

ily)

Ince

ptio

ns

Fam

ily

Affi

liat

edIn

cept

ion

s

Sta

nd-

Alo

ne

min

us

Fam

ily-

Affi

liat

edIn

cept

ion

s

Col

dS

tan

d-A

lon

eM

inu

sH

otS

tan

d-A

lon

eIn

cept

ion

s

Col

dN

oncl

one

Min

us

Hot

Non

clon

eIn

cept

ion

s

Col

dC

lon

eM

inu

sH

otC

lon

eIn

cept

ion

s

Col

dN

oncl

one

Min

us

Hot

Clo

ne

Ince

ptio

ns

Col

dS

tan

d-A

lon

eM

inu

sH

otC

lon

eIn

cept

ion

s

alph

a0.

458*

**0.

233*

**0.

225*

**0.

307*

*0.

350*

*0.

172

0.41

3***

0.55

1***

(8.4

9)(3

.77)

(9.5

1)(2

.34)

(2.5

9)(1

.26)

(3.5

6)(4

.77)

MK

T0.

261*

**0.

265*

**−0

.004

−0.1

14**

*−0

.089

***

−0.0

82**

−0.0

62**

−0.0

33(1

9.65

)(1

7.42

)(−

0.71

)(−

3.54

)(−

2.68

)(−

2.45

)(−

2.17

)(−

1.16

)S

MB

0.15

0***

0.12

9***

0.02

2***

0.04

80.

031

0.00

90.

011

0.09

7***

(9.0

6)(6

.78)

(2.9

6)(1

.19)

(0.7

6)(0

.21)

(0.3

1)(2

.74)

YL

DC

HG

−0.1

11−0

.569

*0.

458*

**−0

.317

−0.7

62−1

.050

−0.6

670.

299

(−0.

40)

(−1.

81)

(3.8

0)(−

0.48

)(−

1.11

)(−

1.51

)(−

1.13

)(0

.51)

BA

AM

TS

Y−1

.779

***

−2.4

67**

*0.

688*

**1.

048

−2.0

83**

−0.5

44−1

.363

*1.

014

(−5.

11)

(−6.

18)

(4.5

1)(1

.24)

(−2.

39)

(−0.

62)

(−1.

82)

(1.3

6)P

TF

SB

D0.

001

−0.0

010.

002

0.00

30.

034*

**0.

039*

**0.

030*

**0.

025*

**(0

.22)

(−0.

29)

(1.2

7)(0

.34)

(3.6

1)(4

.07)

(3.6

8)(3

.09)

PT

FS

FX

0.01

1***

0.01

2***

−0.0

02−0

.010

−0.0

21**

*−0

.019

**−0

.012

*−0

.015

**(3

.51)

(3.4

9)(−

1.11

)(−

1.39

)(−

2.73

)(−

2.37

)(−

1.88

)(−

2.22

)P

TF

SC

OM

0.00

40.

004

0.00

00.

016*

0.00

2−0

.004

−0.0

09−0

.001

(1.0

5)(0

.97)

(−0.

13)

(1.7

1)(0

.19)

(−0.

38)

(−1.

00)

(−0.

08)

Adj

ust

edR

269

.10%

64.3

0%9.

30%

8.60

%8.

40%

6.50

%6.

40%

5.30

%


although it is still significant (t-statistic = 3.77). Stand-alone inceptions out-perform family-affiliated inceptions by as much as 0.225% per month or 2.8%per year (t-statistic = 9.51), supporting the prediction in Hypothesis 3 thatstand-alone inceptions outperform family-affiliated inceptions.

To further understand the impact of inception conditions, we combine fam-ily structure with strategy-based cold and hot inception measures. We createthree groups of inceptions based on their family affiliation: new family incep-tions, family-affiliated nonclone inceptions, and family-affiliated clone incep-tions. Within each group, we further differentiate between cold and hot incep-tions. For each of these six types of inceptions, we form portfolios and regressportfolio returns (as well as pairwise and corner spreads) on hedge fund riskfactors to obtain alpha.

Columns (4) to (6) report the spread between cold and hot inceptions withineach type of family inceptions. Cold inceptions significantly outperform hot in-ceptions when the inceptions are stand-alone (by 0.31% per month) or nonclone(by 0.35% per month). These results support Hypotheses 1 and 4. In contrast,the performance difference between cold and hot clones is insignificant, whichalso supports the prediction in Hypothesis 4, that clone funds are de facto hotand deliver poor performance irrespective of whether they arise in hot or coldconditions.

In the last two columns ((7) and (8)), we examine two groups that synchro-nize the influence of family structure (Hypotheses 3 and 4) and hot/cold incep-tion conditions (Hypothesis 1). Column (7) reports the performance differencebetween family-affiliated nonclone inceptions in cold strategies and clone in-ceptions in hot strategies. The performance spread, 0.41% per month, is posi-tive and significant at the 1% level. The last column reports the difference be-tween cold stand-alone inceptions and hot clone inceptions. The performancedifference is 0.55% per month (6.8% annually), which is not only economicallyand statistically significant, with a t-statistic of 4.77, but also the largest port-folio spread of all those that we report in Table V. Since the corner spreadportfolio (cold stand-alone minus hot clone inceptions) best captures the im-pact of family and demand conditions on inception performance, below we usethis specification to examine the economic source of the performance differencebetween cold and hot inceptions. Figure 2 illustrates the performance differ-ence between cold stand-alone funds and hot clones in event time. The cold-hotspread lasts for at least 10 years, suggesting that strategy demand conditionsand family structure have long-term effects on the performance of new funds.

Panel A of Table VI summarizes the matrix of alpha coefficients for thesesix types of inceptions. The first and second rows report the alphas of in-ceptions launched in cold and hot strategies, respectively. The alpha of thespread portfolio is reported in the third row. Focusing on cold inceptions, we seean interesting pattern: stand-alone (new family) inceptions deliver the high-est alpha, followed by family-affiliated nonclone inceptions and then family-affiliated clone inceptions. A similar pattern emerges for hot inceptions. This


Figure 2. Cumulative abnormal returns after inception by inception type. This figureplots cumulative abnormal returns in event time for cold and hot inception portfolios. We classifya strategy as hot (cold) if its past 36-month returns and flows are among the top (bottom) 30% ofall strategies, and focus on the average postinception returns that can be generated by new stand-alone funds incepted in cold strategies (cold inceptions) and by new clone funds incepted in hotstrategies (hot inceptions).

result further supports our previous conclusion on the performance gap be-tween stand-alone and family-affiliated inceptions.

Tabulated alphas provide additional evidence to support the view that cloneinceptions are de facto hot: clone inceptions in cold strategies generate an al-pha (0.22% per month) that is on par with the alpha of the new family hotinceptions (0.29% per month), suggesting that even clone funds launched incold strategies are likely to encounter an extensive margin that is as high asstand-alone funds launched in hot strategies. To further investigate whetherclone funds are launched to absorb the extra demand for the preceding fund,we compare policies of clone funds and affiliated preceding funds related tofees (both incentive and fixed), redemption notice, and lockup periods. We donot find any significant difference between the policies of clone funds and pre-ceding funds. The average incentive fee for the preceding nonclone funds is15.6%, compared to 15.5% for the follow-up clone funds; the difference is in-significant (t-statistic = 0.896). Thus, clone inceptions do not deviate muchfrom their preceding funds’ policies.

In Panel B, we see that for the corner portfolios (cold nonclone minus hotclone inceptions and cold stand-alone minus hot clone inceptions), the alphasare 0.41% and 0.55% per month (or 5% and 6.8% annually), respectively. Bothalphas are economically large and statistically significant, suggesting thatboth stand-alone and family-affiliated cold inceptions outperform hot cloneinceptions. If cold inceptions are not affiliated with existing families, they


Table VIStrategy Demand and Family Structure: Risk-Adjusted Performance

This table presents a two-way summary of the risk-adjusted returns (alphas) of different typesof inception portfolios. Each portfolio’s alpha is estimated by using the Fung-Hsieh (2004) seven-factor model. We form inception portfolios based on (i) the family structure of each inception (i.e.,the stand-alone inception, or family-affiliated inception including nonclone inceptions and cloneinceptions), and (ii) the strategy-based identification of each inception (i.e., cold or hot inception).In any month, inception portfolios are formed from new hedge fund inceptions of a given familystructure and strategy identification over the prior three months. Each inception is held in its cor-responding portfolio for a 60-month holding period after its inception. The holding period followsthe actual inception date of each fund. Within the holding period, we exclude backfilled returnsand require at least 12 monthly return observations for an inception to be included in any incep-tion portfolio. Funds are equally weighted and rebalanced at the beginning of each month. Theregression equation is presented in Table IV. t-Statistics are in parentheses.

Panel A: Portfolio Alphas by Family Structure and Hot/Cold Strategy Identification

Stand-Alone(New Family) Family-Affiliated Inceptions

Inceptions Nonclone Clone

Cold inceptions 0.600% 0.463% 0.221%(7.446) (4.638) (1.903)

Hot inceptions 0.293% 0.113% 0.050%(2.059) (0.952) (0.504)

Cold minus hot spread 0.307% 0.350% 0.172%(2.344) (2.593) (1.255)

Panel B: Cold-Hot Corner Portfolio Spreads

Cold nonclone minus hot clone spread 0.413%(3.557)

Cold stand-alone minus hot clone spread 0.551%(4.767)

outperform hot clone inceptions by a larger margin than family-affiliated coldinceptions.

Overall, the predictions of our hypotheses are well supported by the data.In later tests, we focus on estimated alphas, and we use the two-way displayof Table VI as a template to analyze performance differences between varioustypes of inceptions.

C. Managerial Experience

A manager with previous experience running a hedge fund may have a pre-existing network of investors from whom to raise capital and has an advantagein the persuading/bargaining step. Since we have identified hot inceptions us-ing only strategy information and family structure, differences in manager ex-perience may confound our measurement. To eliminate effects that managerial


experience may have on our results, we identify and exclude experienced man-agers.

The TASS, HFR, and BarclayHedge databases report the fund managers orprincipals for each fund. After cleaning these names by removing honorificsand similar features (“Ph.D.,” “Doctor,” “Jr.,” etc.), we identify the funds asso-ciated with each manager. Inceptions with a manager associated with a priorfund are marked as having an “experienced” manager. For stand-alone, family-affiliated nonclone, and clone funds, the proportions of funds with experiencedmanagers are 4%, 13%, and 17%, respectively. We remove funds managed byexperienced managers and replicate Table VI using the sample of inexperi-enced managers only.

Table VII presents the results. Overall, the results are similar to those re-ported in Table VI: hot inception portfolios underperform their cold inceptionpeers, particularly among stand-alone inceptions (the spread portfolio’s alphais 0.434% monthly with a t-statistic of 2.83) and nonclone inceptions (alpha =0.336% monthly, t-statistic = 2.39). As reported in the last line of Panel B, theperformance spread between cold stand-alone inceptions and hot clone incep-tions is economically large: 0.723% per month (or 9.0% per year) compared to0.551% per month (or 6.8% per year) in the unrestricted sample (see Table VI).Thus, our results are slightly stronger when using this restricted sample. Byremoving experienced managers who might start a new fund for reasons be-yond our stylized model, such as reputation, the remaining pool of inceptionsmanaged by new managers better fit the setting—and therefore predictions—of our model.

One caveat of this test is that there could be matching errors in our identifi-cation of experienced managers. For instance, since funds only report a snap-shot of fund managers to the databases, some managers may be missed fromthe list of names that we can empirically identify. This could lead to the failureto exclude some experienced managers. For this reason, we include this analy-sis as an additional test, rather than the main specification. Since the removalof experienced managers sharpens our results, if our exclusion methodologyis improved (e.g., due to the availability of more precise data) we expect ourprimary results to be even stronger.

D. Market-Timing and Security-Selection Ability

Thus far, our results are consistent with our model predictions on theimpact of investor strategy demand and family structure on inception per-formance. But what is the nature of the performance difference betweencold and hot inceptions? That is, what kind of managerial skills or strate-gies deliver outperformance? Superior performance could come from gen-uine managerial skills, such as security-selection or market-timing ability.Alternatively, it may be driven by nonstandard risk exposure or it may bean artifact of return-smoothing (e.g., Getmansky, Lo, and Makarov (2004),Cao et al. (2013) (2016)). In this and subsequent sections, we addressthese questions by examining: (i) what skills managers of cold inceptions


Table VIIStrategy Demand and Family Structure: Risk-Adjusted Performance

using the Sample of Inexperienced ManagersThis table presents two-way summary of the risk-adjusted returns (alphas) of different types ofinception portfolios after excluding funds with experienced managers. Each portfolio’s alpha isestimated by using the Fung-Hsieh (2004) seven-factor model. For each fund, we identify the setof principal managers for the fund. If a fund has at least one manager that is associated with afund with an earlier inception date, we mark the fund as having an “experienced” manager. Weexclude these funds from this analysis. Using only funds with “inexperienced” managers, we forminception portfolios based on (1) the family structure of each inception (i.e., the stand-alone incep-tion, or family-affiliated inception including nonclone inceptions and clone inceptions); and (2) thestrategy-based identification of each inception (i.e., cold or hot inception). In any month, inceptionportfolios are formed from new hedge fund inceptions of a given family structure and strategyidentification over the prior three months. Each inception is held in its corresponding portfoliofor a 60-month holding period after its inception. The holding period follows the actual inceptiondate of each fund. Within the holding period, we exclude backfilled returns and require at least12 monthly return observations for an inception to be included in any inception portfolio. Fundsare equally weighted and rebalanced at the beginning of each month. The regression equation ispresented in Table IV. t-Statistics are in parentheses.


Stand-Alone(New Family) Family-Affiliated Inceptions

Inceptions Nonclone Clone

Cold inceptions 0.638% 0.407% 0.177%(7.675) (3.971) (1.427)

Hot inceptions 0.204% 0.071% −0.085%(1.253) (0.571) (−0.771)

Cold minus hot spread 0.434% 0.336% 0.263%(2.825) (2.387) (1.708)


Cold nonclone minus hot clone spread 0.492%(3.889)

Cold stand-alone minus hot clone spread 0.723%(6.029)

possess, (ii) whether there is any difference in performance persistence be-tween cold and hot inceptions, and (iii) how illiquidity and return-smoothingare related to the performance of cold and hot inceptions.

We first examine the security-selection and market-timing skills of cold andhot inceptions. In the literature, the Treynor and Mazuy (1966) model has beenused to evaluate market-timing skills while the Fung and Hsieh (2004) seven-factor model has been used to assess security-selection ability. Here, we usethe Treynor and Mazuy (1966) model augmented with Fung and Hsieh’s sevenrisk factors. Specifically, we estimate the following model separately for the


portfolios of cold stand-alone and hot clone inceptions:

rp,t = αp + βp,1MKTt + γpMKT2t + βp,2SMBt + βp,3YLDCHGt + βp,4BAAMTSYt

+βp,5PTFSBDt + βp,6PTFSFXt + βp,7PTFSCOMt + εp,t, (7)

where for portfolio p the parameters of interest are αp, the selection ability, andγp, the market-timing ability of the fund manager. The coefficient γp measuresmarket-timing skill, that is, how market beta changes, with market conditionforecasts. If a hedge fund manager possesses market-timing ability, she willincrease (decrease) her market exposure before the market goes up (down) andthe timing coefficient γp will be positive.

The abnormal return of portfolio p includes two components: αp and γpM,where M is the long-term mean of MKT2

t . To assess the statistical significanceof the selection coefficient (αp), the timing coefficient (γp), and the abnormal re-turn in the presence of both selection and timing skill (αp + γp × M), we appealto the bootstrap procedure proposed by Kosowski et al. (2006) and Fama andFrench (2010). Details on our bootstrap procedure are outlined in Section IIIof the Internet Appendix. In Table VIII, we present the bootstrapped resultsto evaluate the significance of αp, γp, and (αp + γp × M). Figure 3 graphicallyillustrates the distribution of bootstrapped αp, γp, and (αp + γp × M) and theestimates from our data.

We find that cold inceptions exhibit positive and significant skill in security-selection, but not in market-timing. The selection skill performance (alpha)is 0.54% monthly (6.7% annually) and its bootstrapped p-value is 0, but themarket timing coefficient is insignificant (the bootstrapped p-value is 0.28). Bycontrast, hot inceptions exhibit weaker selection skill (alpha = 0.26% monthly,3.1% annually), and negative (incorrect) and significant market-timing abil-ity, leading to overall insignificant skill-based performance (p-value = 0.54).Cold inceptions significantly outperform hot ones along all three dimensions—selection skill (αp), timing ability (γp), and combined skill (αp + γp × M)—withbootstrapped p-values less than 5%. Hence, it is the superior security-selectionskills possessed by cold-inception managers and incorrect market-timing ofhot-inception managers that drive the performance difference.

An alternative market-timing model is proposed by Henriksson and Mer-ton (1981). To cross-validate our findings, we use the Henriksson and Mer-ton model, augmented with the Fung and Hsieh seven factors, to assess thesecurity-selection and market-timing skills of cold and hot inceptions. We findqualitatively similar results.

E. Performance Persistence

Since performance persistence provides a powerful test of managerial skill,we perform three tests to estimate the degree of performance persistence inhot clone inceptions and cold stand-alone inceptions. A higher degree of per-formance persistence among cold stand-alone inceptions over annual horizons


Table VIIIBootstrapped Security-Selection and Market-Timing Regression

CoefficientsThis table presents the bootstrapped results of security-selection and market-timing analysis forcold stand-alone inceptions and hot clone inceptions. The following security-selection and market-timing regression is applied to each portfolio:

rp,t = αp + βp,1MKTt + γpMKT2t + βp,2SMBt + βp,3YLDCHGt + βp,4BAAMTSYt

+βp,5PTFSBDt + βp,6PTFSFXt + βp,7PTFSCOMt + εp,t ,

where rp,t is the excess return on portfolio p in month t. The independent variables are the marketexcess return (MKT), a size factor (SMB), the monthly change in the 10-year Treasury constantmaturity yield (YLDCHG), the monthly change in Moody’s Baa yield less the 10-year Treasuryconstant maturity yield (BAAMTSY), and three trend-following factors: PFTSBD (bond), PFTSFX(currency), and PFTSCOM (commodity). We also compute the time-series average of MKT2

t , whichwe denote by M. Bootstrapped p-values corresponding to a two-sided test against the null hypoth-esis of (i) no security-selection skill (e.g., α = 0), (ii) no market-timing skill (e.g., γ = 0), and (iii) nosecurity-selection and market-timing skills (e.g., α = 0 and γ = 0 jointly) are reported in squarebrackets.

α γ γ ∗ M α + γ ∗ M

Cold stand-aloneinceptions

0.541% 0.300 0.056% 0.598%

[p-value] [0.000] [0.278] [0.278] [0.000]Hot clone inceptions 0.258% −1.062 −0.199% 0.058%[p-value] [0.023] [0.002] [0.002] [0.541]Cold stand-alone

minus hot clonespread

0.284% 1.362 0.256% 0.540%

[p-value] [0.034] [0.001] [0.001] [0.000]

would suggest that the outperformance in these funds is attributable to man-agerial skill.

We start by dividing the 60-month holding period for each fund into an earlyperiod (1–30 months) and late period (months 31–60). Within each period, wecompute the alpha of each fund using the Fung and Hsieh seven-factor modeland rank funds into quintiles based on their alpha for cold stand-alone andhot clone inceptions. Table IX reports the transition matrix of quintile-basedranks. For instance, Panel A reports that the first-period top-quintile cold andhot inceptions receive a second-period rank of 3.71 and 3.09, respectively,. Ifmanagers have persistent skill, funds ranked high in the first period shouldreceive higher ranks in the second period as well, leading to a positive (cross-period) rank correlation. We can see that cold inceptions exhibit this persis-tence. For cold inceptions, the second-period ranks of funds monotonically in-crease in their first-period ranks with a significant Spearman rank of 30.95%with a p-value of virtually 0. By contrast, hot inceptions exhibit a negative rankcorrelation (significant at the 10% level). We obtain the same conclusion usingthe Spearman test on the ranks directly (without sorting into quintiles). These


Figure 3. Estimated alphas (gammas) versus bootstrapped alpha (gamma) distribu-tions. We plot kernel density estimates of (1) the bootstrapped (α), security skill, distribution,(2) the bootstrapped gamma (γ ), market-timing skill, distribution, and (3) the bootstrapped dis-tribution of (α + γ ∗ M), the combined skill. We present the bootstrapped distribution for threeinception portfolios: the cold stand-alone portfolio, the hot clone portfolio, and the spread portfolio.The vertical lines indicate the parameter estimates from the data.

results suggest that high-performing cold inceptions are likely to continue tohave high performance, while the same cannot be said of high-performing hotinceptions.

Panel B reports results of the performance persistence test proposed byBrown and Goetzmann (1995), which is a nonparametric test based on con-tingency tables. Annual fund returns for cold and hot inceptions are classifiedas winners (W) or losers (L) based on being above or below the median forthat group. Consecutive years for each fund are then classified as Winner-Winner, Winner-Loser, Loser-Winner, or Loser-Loser. Performance persistencefor the group is characterized by more WW and LL than WL and LW. The cross-product ratio (CPR) is the odds ratio of the number of repeat performers to thenumber of those that do not repeat, that is, CPR = (WW × LL)/(WL × LW ).Under the null of no performance persistence, the standardized log(CRP) fol-lows a normal distribution with mean 0 and variance 1. Our sample of cold


Table IXTests of Performance Persistence: Cold versus Hot Inceptions

This table presents results of three performance persistence analyses for hot and cold inceptions.In Panel A, we report the rank persistence over the first 60 months of funds’ lives. For each coldstand-alone and hot clone inceptions, we divide fund performance over the first 60 months intothe early period (months 1 to 30) and the late period (months 31 to 60). Within each period, wecompute the Fung and Hsieh (2004) alpha of each fund. We then rank funds into quintiles basedon their alpha within the period (inceptions with larger alpha values receive high ranks). PanelA reports the average late-period rank for each early-period quintile. For each group, we alsoreport the Spearman rank coefficient of the funds’ early- and late-period ranks (calculated as1 − (6

∑d2

i )/(n3 − n), where di is the difference in quintiles between the early- and late-period forfund i). Panel B reports annual performance persistence using the test from Brown and Goetzmann(1995), who classify annual fund returns as winners or losers and calculate the average logarithmof the odds ratio for cold stand-alone inceptions and hot clone inceptions. Under the null of noperformance persistence, the log of the odds ratio would be 0. The standard error of the log odds

ratio is given by σCPR =√

1WW + 1

LL + 1LW + 1

WL and the Z-statistic is ln(CPR)σCPR

. Panel C reportsresults of the persistence test from Aggarwal and Jorion (2010), who use a one-factor model tocompute annual risk-adjusted returns r∗

i,t = ri,t − β ∗ MKTbt , where MKTb

t is the excess return onthe market portfolio in year t and β is estimated from the first-year monthly returns of each fund.Risk-adjusted returns for either cold stand-alone inceptions or hot clone inceptions over their five-year inception period are then pooled to estimate the AR1 coefficient. The regression equation isr∗{i,t} = α + AR1 × r∗

{i,t−1} + εi,t .

Panel A: Spearman Test of Early/Late Alpha Persistence

Early-period rankCold Inceptions

Late-Period RankHot Inceptions

Late-Period Rank

1 2.21 3.332 2.65 2.973 2.81 2.374 2.95 2.475 3.71 3.09Rank correlation 30.95% −13.88%[p-value] [0.000] [0.055]

Panel B: Brown-Goetzmann Annual Persistence Test

Cold Inceptions Hot Inceptions

Log odds ratio 0.30 −0.10Z-statistic 5.75 −1.67[p-value] [0.000] [0.090]

Panel C: Aggarwal-Jorion Annual Persistence Test

Cold Inceptions Hot Inceptions

AR1 0.167 −0.003[p-value] [0.000] [0.891]


Table XAdditional Evidence from Convertible Arbitrage Funds

This table presents results on how market conditions of convertible bond issuance influence theperformance of inceptions in the convertible arbitrage strategy category. For each inception i in theconvertible arbitrage strategy category, we perform the Fung and Hsieh (2004) seven-factor regres-sion to obtain fund specific alpha, denoted by ai. We then use two variables to describe convertiblebond market conditions: AveNewIssue and AveTotAssets, calculated as the average market cap ofnewly issued convertible bonds and the average total market cap of outstanding convertible bondsin the same 60-month period that we measure the performance of a fund. We then regress fundalpha on three dummy variables, Dcold,i, Dhot,i, and Dothers,i. The dummy variable Dcold,i takes thevalue of 1 if inception i is a cold inception. The other two dummy variables are defined similarly.The regression equation is

αi = b1 × Dcold,i + b2 × Dhot,i + b3 × Dothers,i + b4 × AvgNewIssuei + b5 × AvgTotAssetsi + νi.

Since the coefficients on the market condition variables are very small, we scale AvgNewIssue andAvgTotAssets by 103. This scaling does not affect the significance of our results. t-Statistics are inparentheses. Statistical significance at the 1%, 5%, and 10% levels is denoted by ***, **, and *,respectively.

(1) (2) (3) (4)

DCold 0.638*** 0.534*** 0.899*** 0.774***(6.067) (4.46) (7.31) (6.05)

DHot −0.277 −0.348 0.032 −0.018(−0.92) (−1.15) (0.10) (−0.06)

DOthers 0.216*** 0.115* 0.446*** 0.319***(9.30) (1.89) (7.16) (4.36)

AvgNewIssue (Bn) 14.093* 26.236***(1.81) (3.25)

AvgTotAssets (Bn) −0.337*** −0.427***(−3.98) (−4.82)

Adjusted R2 14.50% 14.80% 16.20% 17.30%Observations 715 715 715 715

stand-alone inceptions has an odds ratio of 0.30 that is positive and signifi-cant (p-value = 0.00), in line with performance persistence among these funds.Hot clone funds, in contrast, have a log odds ratio of −0.10, with an associatedp-value of 0.09. Robustness tests that repeat this experiment using 24-monthreturns yield the same inference, so we relegate these results to the InternetAppendix.

Panel C of Table IX reports results of our third persistence test, based on themethodology proposed by Aggarwal and Jorion (2010). In this test, we examineinception returns in event time over the first 60 months of the fund’s life (e.g.,year 1 runs from the inception month to 12 months later). In each fund-year,we compute the risk-adjusted annual return for that fund from a one-factormodel by subtracting the factor return times the fund’s beta from the fund’sreturn. We then estimate a pooled AR(1) model as follows:

r∗i,t = α + AR1 × r∗

i,t−1 + εi,t, (8)


where r∗i,t is the risk-adjusted return for fund i in year t, and the estima-

tion pools all fund-year risk-adjusted returns for cold stand-alone inceptionsor hot clone inceptions in their five-year inception period. A positive AR1 co-efficient indicates performance persistence at the annual frequency. For coldstand-alone funds, the AR1 coefficient is 0.167, which is statistically signifi-cant with a p-value of essentially 0. Conversely, there is no significant evidenceof performance persistence among our hot clone sample.

In summary, the results of the three tests above indicate that the perfor-mance of cold stand-alone inceptions is persistent over time and that these in-ceptions deliver skill-based performance. In contrast, we do not find evidence ofperformance persistence among hot clone inceptions. Taken together, the testson the managerial skills related to security-selection and market-timing andthe test on performance persistence point to the same conclusion: cold incep-tions deliver outperformance based on genuine managerial skill.

V. Additional Analysis and Alternative Explanations

In this section, we conduct additional analysis to shed light on the nature ofmanagerial skill and to evaluate the robustness of our main findings to alter-native risk factors, fund policies, and empirical specifications.

A. Evidence from CA Funds

We first provide additional evidence on the level of sophistication associatedwith managerial skill using the subsample of CA funds. Convertible bondstend to be underpriced at issue, which provides an arbitrage opportunity forhedge fund managers (see Chan and Chen (2007), Choi et al. (2010, hereafterCGHT)). Thus, the convertible bond market provides an ideal laboratory inwhich to evaluate whether cold CA inceptions benefit mostly from the well-known arbitrage opportunity of bond issuance or whether they derive alphafrom plausibly more sophisticated managerial skills.

We collect convertible bond data from Mergent FISD and SDC from 1989to 2016 and calculate the market cap of newly issued convertible bonds bysumming the dollar value of proceeds in each month. This variable quantifiesthe market-wide arbitrage opportunities associated with new bond issuance.In periods overlapping with CGHT, our variable is very similar to what CGHTplot. We also calculate the total market cap of outstanding convertible bonds toexamine whether the total market size influences hedge fund performance.15

15 One empirical issue in constructing this variable is that we do not observe the conversion de-cisions of all bondholders. Both Chan and Chen (2007) and Choi et al (2010) argue that hedge fundstypically hold convertible bonds for a long period of time because the convergence of underpricedbonds to fundamental value takes a long time. We follow the literature to obtain a reasonableproxy for the total market cap by assuming that all bonds are held until maturity.


Our sample includes 715 CA inceptions. Of these inceptions, 33 are classi-fied as cold and four as hot in our previous analysis.16 Due to this distribution,we focus our analysis on whether cold CA inceptions outperform all other (i.e.,neither hot nor cold) CA inceptions and whether the outperformance origi-nates from the issuance of convertible bonds. For each inception, i, in the CAstrategy, we use the Fung and Hsieh (2004) seven-factor regression to obtainthe fund’s alpha, denoted by ai. We use two variables to describe convertiblebond market conditions: AveNewIssue and AveTotAssets, the average marketcap of newly issued convertible bonds and the average total market cap of out-standing convertible bonds over the fund’s 60-month performance period. Wethen regress alpha on three dummy variables, Dcold,i, Dhot,i, and Dothers,i. Eachdummy takes the value of 1 if i belongs to the corresponding inception group.The regression equation is

αi = b1 × Dcold,i + b2 × Dhot,i + b3 × Dothers,i

+ b4 × AvgNewIssuei + b5 × AvgTotAssetsi + νi. (9)

Table X presents the results of these cross-sectional regressions. In specifi-cation (1) we observe a cold-others spread of 0.42% per month (or, 5.1% an-nually), inferred from the difference between the coefficients on “Cold” and“Others,” with an F-test p-value of 0.00), confirming that cold convertible bondinceptions deliver better performance than other types of inceptions. In spec-ification (2), we include the market cap of newly issued convertible bonds.Convertible bond issuance is positively related to fund alpha, indicating thatCA funds benefit from this well-known arbitrage opportunity. Importantly, thecold-others spread remains largely unchanged (about 0.42%) after controllingfor bond issuance, suggesting that bond issuance conditions do not drive theoutperformance of cold inceptions.17 Specification (3) replaces new issuancewith total market size, which leaves the magnitude of the cold-others spreadalmost unchanged (i.e., about 0.45%). Finally, in specification (4) we control forboth issuance and total market conditions and find that the cold-others spreadremains economically larger and significant, 0.46% monthly (5.7% annually)with an F-test p-value of 0.00.

Although these results focus on a subsector of the hedge fund industry,they provide evidence that cold CA inceptions generate outperformance fromsources other than simple reliance on well-known arbitrage opportunities.

16 The paucity of hot CA inceptions may appear surprising but is reasonable because cold andhot inceptions are defined across strategy categories over time. The performance of the CA strategyhas been relatively smooth during our sample period, and hence the likelihood has been small thatCA would become a category that encounters high investor demand. By contrast, there are periodsin which CA was relatively cold and other categories attracted more capital.

17 Summary statistics further show that cold inceptions on average encounter less issuance ofconvertible bonds than other funds during their performance generating period: the difference inproceeds is about $15.4 billion, with a significant t-statistic of 3.57. If anything, therefore, coldinceptions encounter fewer arbitrage opportunities than other inceptions.


B. Alternative Explanations: Return-Smoothing, Risk Factors, and FundPolicies

To test the robustness of our results, we first examine a leading alternativeexplanation for the performance difference between cold and hot inceptions:exposure to illiquidity or return-smoothing. To address this concern, we followGetmansky, Lo, and Makarov (2004) and Cao et al. (2017) to assess the serialcorrelation associated with illiquidity and return-smoothing. We then comparethe degree of return-smoothing for hot and cold inceptions.

In the interest of space, we summarize our findings here, reporting the for-mal tests in Table IA.IV in the Internet Appendix. We find that both types ofinceptions exhibit return-smoothing. However, hot inceptions have a signifi-cantly higher degree of return-smoothing than cold inceptions do, suggestingthat the returns of hot inceptions benefit more from illiquidity and smoothing.The difference holds not only in summary statistics but also in cross-sectionalanalysis detailed in the Internet Appendix. Hence, the superior performanceof cold over hot inceptions is unlikely to be due to return-smoothing.

Next, although the seven-factor model of Fung and Hsieh (1997) is widelyused in the hedge fund literature to evaluate risk-adjusted returns, funds cannonetheless be exposed to additional risk factors related to liquidity (Pástorand Stambaugh (2003), Sadka (2010)), correlation risk (Buraschi, Kosowski,and Trojani (2014)), economic uncertainty (Bali, Brown, and Caglayan (2014)),and volatility-of-volatility (Agarwal, Arisoy, and Naik (2017)). A study of theinfluence of these factors is important because they can be exploited by hedgefund managers seeking risk-based returns, and we thank the respective au-thors for providing the data. In addition, since nonsynchronous trading of illiq-uid assets can lead to biased estimates of fund beta (see Scholes and Williams(1977)), we include lagged market returns as an additional factor.

In Table IA.V in the Internet Appendix, we start from the seven-factor modelexplaining the alpha of the spread portfolio between cold stand-alone and hotclone inceptions and add additional factors one at a time. Most factors—exceptfor the liquidity factor of Sadka (2010)—do not have significant power in ex-plaining the inception portfolio return spread. Sadka’s liquidity factor is likelyrelevant because it is constructed from variables related to informed trading.We also find that, after the inclusion of the Sadka liquidity factor, our con-clusions from Tables V and VI do not change. The coefficient estimate of risk-adjusted spread (alpha) is close to that reported in column (8) of Table V. Weconclude that our results are not explained by alternative risk factors.

In Table IA.VI in the Internet Appendix, we further examine whether our re-sults can be explained by fund characteristics. We control for fund return char-acteristics, including two measures of risk (market beta, return volatility), twomeasures on how distinctive a fund’s strategy is with respect to other hedgefunds in the same strategy category or with respect to common risk factors inthe market (the Strategy Distinctiveness Index [SDI] of Sun, Wang, and Zheng(2012) and the R2 of Amihud and Goyenko (2013)), as well as operational pol-icy choices related to incentive fees, management fees, the redemption notice


period, and redemption frequency (we express redemption frequency in days—a larger value indicates a more restrictive redemption policy).

To achieve this goal, we conduct a cross-sectional analysis by linking fund-specific alpha to a dummy variable indicating the type of an inception (i.e.,whether is it a cold stand-alone inception) as well as fund characteristics. TheInternet Appendix provides more details of the analysis. Our main finding isthat the cold stand-alone dummy is associated with a significant and positivecoefficient in this cross-sectional regression, and that fund characteristics donot absorb this significance. For instance, although SDI is positively related tofund alpha, our main result remains unchanged (cold stand-alone inceptionsare associated with higher alphas). Controlling for other characteristics andpolicies yields a very similar result.

Taken together, our findings suggest that the outperformance of cold incep-tions cannot be attributed to exposure to return-smoothing, additional risk fac-tors, characteristics of fund returns, fund policy choices, or flows. These resultslend further support to our conclusion that cold inceptions deliver performancebecause managers of these inceptions possess genuine skill.

C. Alternative Holding Periods and Definitions of Cold and Hot inceptions

Next, we use alternative holding periods and alternative definitions of coldand hot inceptions to investigate their impact on our main findings. Since thereis a trade-off between the length of a holding period and the number of fundsavailable, our baseline analysis adopts a holding period of 60 months, whichis often used in the literature to estimate portfolios’ dynamic risk exposure.We re-do Table VI using a shorter holding period of 48 months (Panels A1 andB1 of Table IA.VII in the Internet Appendix) and a longer period of 72 months(Panels A2 and B2 of Table IA.VII in the Internet Appendix). We see that themain features of Table VI remain unchanged. It is perhaps not surprising tosee the robustness of our results over different testing horizons considering ourprevious finding that cold inceptions are associated with persistent managerialskills that are superior to hot inceptions.

We also examine the impact of alternative definitions of cold and hot incep-tions. In our main analysis, inceptions are classified as hot when they investin strategies with high investor demand as proxied by high strategy categoryreturn and flows. Since (past) category inceptions provide another observablesignal of investor demand, we can also define hot inceptions as those investedin strategies with high recent strategy category inceptions. A strategy is clas-sified as hot (cold) if its normalized inceptions are among the top (bottom) 30%of all strategies over the 36 months prior to inception. Inceptions are normal-ized by dividing by the number of funds in that strategy at the beginningof the given period. We form inception portfolios based on both family struc-ture and this alternative strategy identification. The results are reported inTable IA.VIII in the Internet Appendix. Our main conclusions from Table VIremain largely unchanged.


D. The Impact of Data Biases

The seminal papers of Fung and Hsieh (1997, 2000) and Jorion and Schwarz(2019) document biases (e.g., survivorship bias, backfill bias, and selectionbias) in hedge fund data and their impact on hedge fund performance. Here,we address concerns related to these biases. The Internet Appendix providesadditional discussion regarding the bias related to voluntary reporting todatabases.

To mitigate survivorship bias, we include both live and defunct funds fromeach of the three databases. To minimize measurement errors caused by fundsreporting after they have been alive for some time, we measure inception pe-riods from the true inception date (not the date at which the fund starts re-porting to the database nor the date of the first available return for the fund).As mentioned in Sections II and IV, we address the backfill bias (caused byfunds choosing to backfill their returns only if they are proud of their earlyperformance) using the method developed in Jorion and Schwarz (2019). Thisinvolves estimating the date at which each fund is added to the database usinginformation from the cross-section of fund IDs for each database, then mark-ing returns before the add date for each fund as missing. Jorion and Schwarz(2019) show that this is a more effective method of eliminating the backfillbias than deleting early performance (12 or 24 months) from all funds, as isfrequently done in the literature.

VI. Conclusion

In this paper, we explore the economics of hedge fund inceptions in the pres-ence of search frictions. To do so, we incorporate into the Berk and Green (2004)model one of the most important types of frictions in the hedge fund industry,namely, new managers’ need to search for accredited investors. The novel in-tuition from our stylized model is that investor demand and performance in-fluence the search-and-bargaining process associated with raising capital fornew funds. The substitution effect between the extensive and intensive mar-gins of capital-raising gives rise to two different types of inceptions: hot in-ceptions that replicate the strategy of existing funds and cold inceptions thatdeliver new skills and superior performance. Moreover, family structure arisesendogenously to reduce the search frictions, but negatively affects the perfor-mance incentives of affiliated nonclone inceptions. Since search frictions am-plify diseconomies of scale, they also motivate the inception of clone funds.

Empirically, we develop proxies for strategy popularity among investors andfor family structure. We find that funds arising in strategies with high investordemand (i.e., hot inceptions) subsequently underperform those facing demandheadwind at inception (i.e., cold inceptions) on a risk-adjusted basis. We fur-ther find that cold inceptions, but not hot inceptions, outperform existing fundsand that family-affiliated inceptions underperform stand-alone inceptions.Using an identification involving both strategy popularity and family struc-ture, we sharpen our results by identifying cold stand-alone inceptions andhot clone inceptions that exhibit a risk-adjusted performance spread as high as


6.8% per year. We further show that the performance difference is attributableto genuine managerial skills that managers of cold inceptions bring into thehedge fund industry, as opposed to loading on alternative sources of risk suchas illiquidity or return-smoothing. Cold inceptions also exhibit significant per-formance persistence, suggesting that their performance is skill-based. Finally,tests excluding experienced managers lead to stronger results with larger eco-nomic performance differences between cold stand-alone and hot clone incep-tions.

Overall, our findings suggest that market frictions are an important eco-nomic mechanism that drives cross-sectional variation in the risk-adjustedperformance of hedge fund inceptions and leads to the formation of familystructure in the industry. Importantly, we show that it is possible to distin-guish ex ante new funds that provide genuine innovations to the industry. Ourmodel, methodology, and empirical analysis have important normative impli-cations that may also apply to other fast-growing markets in which managersneed to actively search for capital. Private equity funds and private pensionfunds are two examples. Although our analysis focuses on the cross-section ofhedge fund strategies, the search mechanism may also impact time-series pat-terns of fund returns. Our results call for more attention to market frictionsthat new funds face as we seek to better understand the incentives and overallvalue of the hedge fund industry.

Initial submission: October 25, 2017; Accepted: April 30, 2020Editors: Stefan Nagel, Philip Bond, Amit Seru, and Wei Xiong

REFERENCES

Ackermann, Carl, Richard McEnally, and David Ravenscraft, 1999, The performance of hedgefunds: Risk, return, and incentives, Journal of Finance 54, 833–874.

Agarwal, Vikas, and Naveen D. Naik, 2004, Risks and portfolio decisions involving hedge funds,Review of Financial Studies 17, 63–98.

Agarwal, Vikas, Naveen D. Daniel, and Narayan Y. Naik, 2009, Role of managerial incentives anddiscretion in hedge fund performance, Journal of Finance 64, 2221–2256.

Agarwal, Vikas, Naveen D. Daniel, and Narayan Y. Naik, 2011, Do hedge funds manage theirreported returns?, Review of Financial Studies 24, 3281–3320.

Agarwal, Vikas, Y. Eser Arisoy, and Narayan Y. Naik, 2017, Volatility of aggregate volatility andhedge fund returns, Journal of Financial Economics 125, 491–510.

Aggarwal, Rajesh K., and Philippe Jorion, 2010, The performance of emerging hedge funds andmanagers, Journal of Financial Economics 96, 238–256.

Amihud, Yakov, and Ruslan Goyenko, 2013, Mutual fund’s R2 as predictor of performance, Reviewof Financial Studies 26, 667–694.

Aragon, George O., and Vikram Nanda, 2012, Tournament behavior in hedge funds: High-watermarks, fund liquidation, and managerial stake, Review of Financial Studies 25, 937–974.

Bali, Turan G., Stephen J. Brown, and Mustafa O. Caglayan, 2014, Macroeconomic risk and hedgefund returns, Journal of Financial Economics 114, 1–19.

Berk, Jonathan B., and Richard C. Green, 2004, Mutual fund flows and performance in rationalmarkets, Journal of Political Economy 112, 1269–1295.

Berk, Jonathan B., and Jules H. van Binsbergen, 2015, Measuring skill in the mutual fund indus-try, Journal of Financial Economics 118, 1–20.

Berk, Jonathan B., and Jules H. van Binsbergen, 2017, Mutual funds in equilibrium, AnnualReview of Financial Economics 9, 147–167.

http://dx.doi.org/10.1111/0022-1082.00129

http://dx.doi.org/10.1093/rfs/hhg044

http://dx.doi.org/10.1111/j.1540-6261.2009.01499.x

http://dx.doi.org/10.1093/rfs/hhr058

http://dx.doi.org/10.1016/j.jfineco.2017.06.015


http://dx.doi.org/10.1093/rfs/hhs182

http://dx.doi.org/10.1093/rfs/hhs182



http://dx.doi.org/10.1086/424739


http://dx.doi.org/10.1146/annurev-financial-110716-032454

http://dx.doi.org/10.1146/annurev-financial-110716-032454


Berk, Jonathan B., Jules H. van Binsbergen, and Binying Liu, 2017, Matching capital and labor,Journal of Finance 72, 2467–2504.

Bhattacharya, Utpal, Jung H. Lee, and Veronika K. Pool, 2013, Conflicting family values in mutualfund families, Journal of Finance 68, 173–200.

Brown, Stephen J., and Wukkuan B. Geotzmann, 1995, Performance persistence, Journal of Fi-nance 50, 679–698.

Buraschi, Andrea, Robert Kosowski, and Fabio Trojani, 2014, When there is no place to hide:Correlation risk and the cross-section of hedge fund returns, Review of Financial Studies 27,581–616.

Cao, Charles, Yong Chen, Bing Liang, and Abdrew Lo, 2013, Can hedge funds time market liquid-ity? Journal of Financial Economics 109, 493–516.

Cao, Charles, Grant Farnsworth, Bing Liang, and Andrew W. Lo, 2017, Return-smoothing, liquid-ity costs, and investor flows: Evidence from a separate account platform, Management Science63, 2233–2250.

Cao, Charles, Bradley A. Goldie, Bing Liang, and Lubomir Petrasek, 2016, What is the nature ofhedge fund manager skills? Evidence from the risk arbitrage strategy, Journal of Financialand Quantitative Analysis 51, 929–957.

Carhart, Mark M., 1997, On persistence in mutual fund performance, Journal of Finance 52, 57–82.

Chan, Alex W. H., and Nai-fu Chen, 2007, Convertible bond underpricing: Renegotiable covenants,seasoning, and convergence, Management Science 53, 1793–1814.

Choi, Darwin, Mila Getmansky, Brian Henderson, and Heather Tookes, 2010, Convertible bondarbitrageurs as suppliers of capital. Review of Financial Studies 23, 2492–2522.

Choi, James J., David Laibson, and Brigitte C. Madrian, 2010, Why does the law of one price fail?An experiment on index mutual funds, Review of Financial Studies 23, 1405–1432.

Duffie, Darrell, 2010, Presidential address: Asset price dynamics with slow-moving capital,Journal of Finance 65, 1237–1267.

Duffie, Darrell, Nicolae Gârleanu, and Lasse Heje Pedersen, 2005, Over-the-counter markets,Econometrica 73, 1815–1847.

Duffie, Darrell, Nicolae Gârleanu, and Lasse Heje Pedersen, 2007, Valuation in over-the-countermarkets. Review of Financial Studies 20, 1865–1900.

Fama, Eugene F., and Kenneth French, 2010, Luck versus skill in the cross-section of mutual fundreturns, Journal of Finance 65, 1915–1947.

Fang, Jieyan, Alexander Kempf, and Monika Trapp, 2014, Fund manager allocation, Journal ofFinancial Economics 111, 661–674.

Fung, William, and David A. Hsieh, 1997, Empirical characteristics of dynamic trading strategies:The case of hedge funds. Review of Financial Studies 10, 275–302.

Fung, William, and David A. Hsieh, 2000, Performance characteristics of hedge funds and com-modity funds: Natural vs. spurious biases. Journal of Financial and Quantitative Analysis 35,291–307.

Fung, William, and David A. Hsieh, 2001, The risk of hedge fund strategies: Theory and evidencefrom trend followers, Review of Financial Studies 14, 313–341.

Fung, William, and David A. Hsieh, 2004, Hedge fund benchmarks: A risk-based approach,Financial Analysts Journal 60, 65–80.

Fung, William, David A. Hsieh, Narayan Y. Naik, and Tarun Ramadorai, 2008, Hedge funds: Per-formance, risk, and capital formation, Journal of Finance 63, 1777–1803.

Gârleanu, Nicolae, and Lasse Heje Pedersen, 2018, Efficiently inefficient markets for assets andasset management, Journal of Finance 73, 1663–1712.

Getmansky, Mila, Andrew W. Lo, and Igor Makarov, 2004, An econometric model of serial correla-tion and illiquidity in hedge fund returns, Journal of Financial Economics 74, 529–609.

Glode, Vincent, and Richard C. Green, 2011, Information spillovers and performance persistencefor hedge funds, Journal of Financial Economics 101, 1–17.

Goetzmann, William N., Jonathan E. Ingersoll Jr., and Stephen Ross, 2003, High-water marks andhedge fund management contracts, Journal of Finance 58, 1685–1718.

http://dx.doi.org/10.1111/jofi.12542

http://dx.doi.org/10.1111/j.1540-6261.2012.01797.x

http://dx.doi.org/10.1111/j.1540-6261.1995.tb04800.x


http://dx.doi.org/10.1093/rfs/hht070


http://dx.doi.org/10.1287/mnsc.2015.2401

http://dx.doi.org/10.1017/S0022109016000387

http://dx.doi.org/10.1017/S0022109016000387


http://dx.doi.org/10.1287/mnsc.1070.0722

http://dx.doi.org/10.1093/rfs/hhq003

http://dx.doi.org/10.1093/rfs/hhp097

http://dx.doi.org/10.1111/j.1540-6261.2010.01569.x

http://dx.doi.org/10.1111/j.1468-0262.2005.00639.x

http://dx.doi.org/10.1093/rfs/hhm037

http://dx.doi.org/10.1111/j.1540-6261.2010.01598.x



http://dx.doi.org/10.1093/rfs/10.2.275

http://dx.doi.org/10.1093/rfs/14.2.313

http://dx.doi.org/10.2469/faj.v60.n5.2657

http://dx.doi.org/10.1111/j.1540-6261.2008.01374.x



http://dx.doi.org/10.1111/1540-6261.00581


Grossman, Sanford J., and Joseph E. Stiglitz, 1980, On the impossibility of informationally effi-cient markets, American Economic Review 70, 393–408.

Henriksson, Roy D., and Robert C. Merton, 1981, On market timing and investment performance:Statistical procedures for evaluating performance skills, Journal of Business 54, 513–533.

Hortaçsu, Ali, and Chad Syverson, 2004, Product differentiation, search costs and competitionin the mutual fund industry: A case study of S&P 500 index funds, Quarterly Journal ofEconomics 119, 303–356.

Jackson, Matthew O., and Brian W. Rogers, 2007, Meeting strangers and friends of friends: Howrandom are social networks? American Economic Review 97, 890–915.

Jegadeesh, Narasimhan, and Sheridan Titman, 1993, Returns to buying winners and sellinglosers: Implications for stock market efficiency, Journal of Finance 48, 65–91.

Jiao, Yawen, Massimo Massa, and Hong Zhang, 2016, Short selling meets hedge fund 13F: Ananatomy of informed demand, Journal of Financial Economics 122, 544–567.

Jorion, Philippe, and Christopher Schwarz, 2019, The fix is in: Properly backing out backfill bias,Review of Financial Studies 32, 5048–5099.

Jylhä, Petri, and Matti Suominen, 2011, Speculative capital and currency carry trades, Journal ofFinancial Economics 99, 60–75.

Kosowski, Robert, Narayan Naik, and Melvyn Teo, 2007, Do hedge funds deliver alpha? ABayesian and bootstrap analysis. Journal of Financial Economics 84, 229–264.

Kosowski, Robert, Allan Timmermann, Russ Wermers, and Hal White, 2006, Can mutual fund“stars” really pick stocks? New evidence from a bootstrap analysis. Journal of Finance 61,2551–2595.

Massa, Massimo, 2003, How do family strategies affect fund performance? When performancemaximization is not the only game in town, Journal of Financial Economics 67, 249–305.

McMillan, John, and Michael Rothschild, 1994, Search, 905–927, in R. J. Aumann and S. Hart,eds.: Handbook of Game Theory, Vol.2 (Elsevier Science, Amsterdam).

Nanda, Vikram, Z. Jay Wang, and Lu Zheng, 2004, Family values and the star phenomenon:Strategies of mutual fund families, Review of Financial Studies 17, 667–698.

Pástor, Lubos, and Robert F. Stambaugh, 2003, Liquidity risk and expected stock returns, Journalof Political Economy 111, 642–685.

Rubinstein, Ariel, and Asher Wolinsky, 1985, Equilibrium in a market with sequential bargaining,Econometrica 53, 1133–1150.

Sadka, Ronnie, 2010, Liquidity risk and the cross-section of hedge-fund returns, Journal of Finan-cial Economics 98, 54–71.

Scholes, Myron, and Joseph Williams, 1977, Estimating betas from nonsynchronous data, Journalof Financial Economics 5, 309–327.

Sirri, Erik R., and Peter Tufano, 1998, Costly search and mutual fund flows, Journal of Finance53, 1589–1622.

Sun, Lin, Zheng Sun, and Lu Zheng, 2020, When the start matters: Investor sentiment andhedge fund inceptions, Working paper, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3231259.

Sun, Zheng, Ashley Wang, and Lu Zheng, 2012, The road less traveled: Strategy distinctivenessand hedge fund performance, Review of Financial Studies 25, 96–143.

Treynor, Jack L., and Kay K. Mazuy, 1966, Can mutual funds outguess the market? HarvardBusiness Review 44, 131–136.

Supporting Information

Additional Supporting Information may be found in the online version of thisarticle at the publisher’s website:

Appendix S1: Internet Appendix.Replication Code.

http://dx.doi.org/10.1086/296144

http://dx.doi.org/10.1257/aer.97.3.890



http://dx.doi.org/10.1093/rfs/hhz024


http://dx.doi.org/10.1111/j.1540-6261.2006.01015.x

http://dx.doi.org/10.1016/S0304-405X(02)00253-2

http://dx.doi.org/10.1093/rfs/hhg054

http://dx.doi.org/10.2307/1911015



http://dx.doi.org/10.1016/0304-405X(77)90041-1

http://dx.doi.org/10.1016/0304-405X(77)90041-1

http://dx.doi.org/10.1111/0022-1082.00066

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3231259

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3231259


Internet Appendix: Page 1

Internet Appendix for “The Economics of Hedge Fund Startups:

Theory and Empirical Evidence”

CHARLES CAO, GRANT FARNSWORTH, and HONG ZHANG1

This document provides supplementary material to the paper “The Economics of Hedge Fund Startups: Theory and Empirical Evidence.” Section I presents all of the model proofs as well as the parameter values for the baseline case examined in our numerical example (Table IAI). Section II presents an investor-search model and a comparison with the manager-search model presented in the main text. Section III provides more details on the three databases (Table IAII) and strategy-matching (Table IAIII). Section IV presents details on the market-timing test. Finally, Section V details robustness checks discussed in the main test.

1 DOI String


I. Proofs of the Manager-Search Model

A. Proof of Proposition 1

We first derive the first order conditions (FOC) for a new fund manager. To do so, we notice that the

generation and distribution of extra performance do not change equation (1): if the manager generates

𝜙 + δ for period 𝑡 and delivers δ to investors, we have 𝐸(𝑟 ) = 𝜙 + δ − 𝑐(𝑞 ) − 𝑠(𝑞 ) = δ and thus

the same Berk and Green (2004, BG) condition of 𝜙 − 𝑐(𝑞 ) − 𝑠(𝑞 ) = 0. Plugging this condition into

the manager’s utility function, we get 𝑈(δ) = 𝑓 × 𝑞 − 𝐿 δ = × (𝜙 − 𝑓 − 𝑠(𝑡)) − 𝐿 δ. The FOC leads

to 𝑈′ (δ) = − 𝑠′ − 𝐿 = 0 and thus 𝐿 = − 𝑠′ .

Next, it would be suboptimal for the manager to pay additional search costs if she has already raised

enough capital to match the optimal size of the fund. Hence, the two sides of condition ii equalize in

equilibrium, and the search cost in (1) can be quantified as 𝑠(𝑡) =( )

( )=

× ( )=

( )[ ( ) ( ) ( )].

From this optimal search cost, we get 𝑠′ = −( ) ′

( ) ( ) ( ) ( ) as its derivative with respect to δ. We

note that 𝑠′ < 0, so performance can reduce effective search costs. Plugging 𝑠′ into the FOC, we have

𝐿 = ∙( )

( ) ( ) ( ) ( ). Further, notice that 𝜉(δ) = 𝜉 × δ, so we can solve for the optimal value

of δ as 𝑓 (𝐿 𝑏𝜇 (𝑡)𝜉 𝑎𝑧 ) −( )

( ). Since δ cannot take negative values because the new fund cannot

short the flow of 𝑙𝑛-investors, we have its optimal value, 𝛿∗ = 𝑀𝑎𝑥{0, 𝑓 (𝐿 𝑏𝜇 (𝑡)𝜉 𝑎𝑧 ) −( )

( )},

which proves property i. The proof of property ii is straightforward using this closed-form solution of 𝛿∗.

Hence, encountering high investor demand reduces the optimal alpha. We also plug 𝛿∗ into the BG

condition to solve for the optimal fund size, 𝑞∗(𝑡) = −( )

for nonzero 𝛿∗, which is

essentially the standard BG fund size, , with a search friction-initiated adjustment. Note that search

frictions reduce the optimal size that the fund can manage.

To prove property iii, we start from the δ∗ that satisfies the FOC conditioned on the value of 𝑧 :

𝑈′ (δ∗, 𝑧 ) =( , )

| ∗ = 0. Next, assume that 𝑧 has experienced a small change, 𝑑𝑧, leading the

manager to adjust δ∗ by 𝑑δ∗. For the change in 𝑑δ∗ to be optimal, we have 𝑈′ (δ∗ + 𝑑δ∗, 𝑧 + 𝑑𝑧) = 0,

which requires the first-order impact of 𝑑𝑧 on 𝑈′ to be offset by that of 𝑑δ∗, or ′ ( ∗, )

× 𝑑𝑧 +


′ ( ∗, )

∗ × 𝑑δ∗ = 0. Hence, the influence of 𝑑𝑧 on optimal alpha can be written as ∗

=

−′ ( ∗, )

/′ ( ∗, )

∗ .

To quantify the sign of this influence, we notice that 𝑈′ (δ∗, 𝑧 ) = − 𝑠′ − 𝐿′ , which is similar to the

FOC that we derived above, except that we allow for a potential convex 𝐿(δ). Because 𝐿(δ) is independent

of 𝑧 , we have ′ ( ∗, )

= −′

. Since 𝑠′ = −( ) ′

( ) ( ) ( ) ( ), we have

′

> 0 and ′ ( ∗, )

<

0. Likewise, ′ ( ∗, )

∗ = −′

−′

= − 𝑠′′ − 𝐿′′ . Since 𝑠′′ > 0 (and thus − 𝑠′′ < 0), for any convex

or linear learning cost 𝐿′′ ≥ 0, we have ′ ( ∗, )

∗ = − 𝑠′′ − 𝐿′′ < 0. Overall, we have ∗

=

−′ ( ∗, )

/′ ( ∗, )

∗ < 0, which shows that δ∗ decreases in 𝑧(𝑡) under convex learning costs. Hence, we

prove property iii: a convex learning cost does not affect our results. Note that the intermediate condition, ′ ( ∗, )

∗ = 𝑈′′ (δ∗, 𝑧 ) < 0, also confirms that δ∗ maximizes the manager’s utility.

To add intuition, below we show that the benefit of learning is concave in alpha. Notice that a small

fund-specific-alpha (𝑑δ) above zero will relax the BG condition from 𝜙 − 𝑏(𝑞 ) − 𝑓 − 1/

[𝑎𝑧(𝑡)𝜇 (𝑡)] = 0 to 𝜙 − 𝑏(𝑞 + 𝑑𝑞 ) − 𝑓 − 1/{𝑎𝑧(𝑡)[𝜇 (𝑡) + 𝜇 (𝑡) 𝜉(𝑑𝛿)]} = 0, where 𝑑𝑞 refers

to enhanced equilibrium fund size due to 𝑑δ, or the benefit of generating 𝑑δ. Comparing the two conditions,

we have 𝑑𝑞 = {−1/{𝑎𝑧(𝑡)[𝜇 (𝑡) + 𝜇 (𝑡) 𝜉(𝑑𝛿)]} + 1/[𝑎𝑧(𝑡)𝜇 (𝑡)]}. We see that 𝑑𝑞 is an

increasing and concave function of 𝑑δ.

Two more properties are useful for further analysis. First, following the above method, we have ∗

=

−′ ( ∗, )

/′ ( ∗, )

∗ < 0, because ′ ( ∗, )

= − 𝑠′ − 𝐿′ = −′

< 0. Hence, δ∗ also

decreases in 𝜇 with convex learning costs. This conclusion is not needed for Proposition 1 but will be

used in Proposition 2 when we study how family affiliation affects 𝛿∗. This property suggests that the

negative influence of family affiliation (which enhances 𝜇 (𝑡)) on inception performance holds with a

convex function of learning cost.

Second, although we have used the cost function, 𝑐(𝑞 ) = 𝑏𝑞 + 𝑓, to prove the first property, the

FOC can be applied to more general cost functions as long as ( )

= 𝑏. To see this, we derive from

equation (3) the FOC for optimal δ as 𝐿 = 𝑓 × . To calculate , we examine how additional

performance, 𝑑δ, influences the optimal fund size under the BG condition by taking the derivative of the

BG condition with respect to δ, that is, (𝜙 − 𝑐(𝑞 ) − 𝑠(𝑡)) = 0. Since = 0 and ( )

=( )

,


we have −( )

−( )

= 0, or = − 𝑠′ , which we refer to as the marginal BG condition. This

result is intuitive: better performance translates into larger fund size (i.e., ) when a reduced search cost

enables a larger amount of capital (i.e., − 𝑠′ ). Plugging this condition into 𝐿 = 𝑓 × leads to the same

FOC.

B. Lemma IA1 and its Proof:

LEMMA IA1: When a new fund manager retains a fraction of fund-specific alpha, to pay for the operation

costs of the fund or as incentive fees, 𝛿∗ decreases in 𝑧(𝑡).

Proof: We first show that distributing only a fraction of fund-specific alpha to investors—and retaining the

balance to enhance 𝜙 in the BG condition, offsetting the operation or search costs of the fund—does not

affect property ii of Proposition 1. Mathematically, the BG condition can be written as 𝜙 + 𝛿 − 𝑐(𝑞 ) −

𝑠(𝑡) = 𝑚𝛿 in this case, where 𝑚 denotes the fraction of 𝛿 distributed to investors. Rewrite the BG condition

as 𝜙 + 𝛿 ′ − 𝑐(𝑞 ) − 𝑠(𝑡) = 0, where 𝛿 ′ = (1 − 𝑚)𝛿 refers to the fraction of fund alpha that can be used

by the fund to pay its operational and search costs. Meanwhile, since low-type investors get only 𝑚𝛿, their

capital flow reduces proportionately. Hence, the new manager can raise capital of 𝑞(𝑡) =

𝜌 (𝜇 (𝑡) + 𝜇 (𝑡)𝑚𝜉 δ)𝑧(𝑡) = 𝜌 𝜇 (𝑡) + 𝜇 (𝑡)𝜉′ δ′ 𝑧(𝑡), where 𝜉′ = × 𝜉 . In this case, we

rewrite search cost and utility as 𝑠(𝑡) =( )[ ( ) ( ) ′ ′]

and 𝑈(δ) = 𝑓 × 𝑞 − 𝐿 δ = 𝑓 × 𝑞 −

𝐿′ δ′ = 𝑈(δ′), where 𝐿′ = 𝐿 /(1 − 𝑚). Therefore, the new manager maximizes 𝑈(δ′) = 𝑓 × 𝑞 − 𝐿′ δ′ by

adjusting δ′, subject to the BG condition, 𝜙 + 𝛿 ′ − 𝑐(𝑞 ) − 𝑠(𝑡) = 0, and the search condition, 𝑠(𝑡) =

( )[ ( ) ( ) ′ ′].

Plugging 𝑞 = (𝜙 + 𝛿 ′ − 𝑓 − 𝑠(𝑡)) into 𝑈(δ′) and taking the derivative of 𝑈(δ′) w.r.t. δ′, we have

𝑈 ′′ (δ′) = 1 −

( )′ − 𝐿′ = 0, or 𝐿′ = ∙ 1 +

( ) ′

( ) ( ) ( ) ′ ′ as the FOC, which allows us to

solve for the optimal value of δ′ as 𝛿 ′∗ = 𝑀𝑎𝑥{0, 𝑎𝑧′

− 1 𝜇 (𝑡)𝜉′ −( )

( ) ′ }. The

corresponding optimal fund size for nonzero 𝛿 ′∗ is 𝑞∗(𝑡) = − [′ /

′ / ( ) ′+

( )

( ) ′ ].

Thus, 𝛿 ′∗ decreases in 𝑧 . Since 𝛿 ′ = (1 − 𝑚)𝛿, 𝛿∗ decreases in 𝑧 as well.

In addition, we see that 𝛿∗ decreases in 𝜇 (𝑡) under this retaining policy. This property is important

when studying the influence of the family structure in Proposition 2. In a nutshell, it confirms that the above


retaining policy will not affect the main conclusion of Proposition 2 that family affiliation, which enhances

𝜇 (𝑡), has a negative influence on 𝛿∗.

We now examine the case in which the new fund manager retains as incentive fees a fraction, 𝜆, of

before-fee performance. In this case, the manager maximizes the modified payoff function 𝑈(𝛿) = (𝑓 +

𝜆𝛿)𝑞 − 𝐿(𝛿), subject to the same search condition of (2) and the modified BG condition 𝐸(𝑟 ) = 𝜙 +

(1 − 𝜆)𝛿 − 𝑏𝑞 − 𝑓 − 𝑠(𝑡) = (1 − 𝜆)𝛿, or simply 𝜙 − 𝑏𝑞 − 𝑓 − 𝑠(𝑡) = 0. But since 𝑙𝑛-investors only

(1 − 𝜆)𝛿, their capital demand reduces proportionally. Hence, the new manager can raise capital 𝑞(𝑡) =

𝜌 (𝜇 (𝑡) + 𝜇 (𝑡)(1 − 𝜆)𝜉 δ)𝑧(𝑡). The search cost becomes 𝑠(𝑡) =( )[ ( ) ( ) ]

.

To derive the optimal 𝛿, we first plug the BG condition into the manager’s utility function and get

𝑈(𝛿) = (𝑓 + 𝜆𝛿)( )

− 𝐿 𝛿. We then take the first-order derivative of 𝑈(𝛿) w.r.t. 𝛿, which leads to

the FOC of ( )

= −( )

− (( )

+( )

) + (𝜙 − 𝑓) − 𝐿 = 0. Plugging ( )

=

−( )( )

( ( ) ( )( ) ) ( ) into the FOC, we get

( )( )

( ( ) ( )( ) ) ( )−

( )

( ( ) ( )( ) ) ( )+ (𝜙 − 𝑓) − 𝐿 = 0, or (𝜇 (𝑡) + 𝜇 (𝑡)(1 − 𝜆)𝜉 δ) =

( )( ) ( )

( ) ( ). Hence, we solve for the optimal value of δ as δ∗ =

𝑀𝑎𝑥{0,( )

[( )

( )

( )

( ) ( ) ( )−

( )

( )]}. When a valid and nonnegative interior optimal

solution exists (i.e., δ∗ > 0, which also requires 𝑏𝐿 − 𝜆(𝜙 − 𝑓) > 0), we can easily see that optimal fund

alpha, δ∗, decreases in both investor demand 𝑧(𝑡) and the mass of high-type noninvestors 𝜇 (𝑡) that the

manager encounters in the search step.

C. Lemma IA2 and its Proof:

LEMMA IA2: A steady-state of old funds exists in which these funds optimally choose not to pursue

additional alpha, whereas the behavior of new funds follows Proposition 1.

Proof: We now examine the search process and incentives of an existing benchmark fund. As mentioned,

a general fund-raising procedure also occurs for the existing fund at the beginning of period 𝑡, in which

(some) old investors withdraw and the fund searches for new investors. Once the investor clientele of the

fund is stabilized (i.e., in the steady-state), investors attracted by the fund will contribute capital, which is

then locked up during period 𝑡 for the fund to manage and invest. At the end of the period, investors receive

search-cost-adjusted payoffs. Only after the existing fund stabilizes and locks in its capital for investment

will the new fund start to raise capital. This sequence of moves is reasonable in practice (i.e., the existing


pool of investors allows the existing fund to draw capital first) and allows the optimal policies of the new

funds to be calibrated against the steady-state of the existing fund in our stylized model.

The fund-raising procedure of existing funds follows a similar search-and-bargaining process as the

new fund. In particular, the benchmark fund also searches for the remaining investors in the market (i.e.,

their noninvestors): by paying a search cost 𝑇𝑆 , noninvestors will be matched with the fund at a speed of

𝜌 = 𝑎 × 𝑇𝑆 per unit of time where the superscript “E” denotes the existing fund. By delivering an

expected extra performance above 𝜙 , which we also refer to as δ when there is no risk of confusion, the

existing fund can persuade 𝑙𝑛-investors to invest with a probability 𝜉(δ).

An existing fund already has investors for whom it does not need to search. Its steady-state, therefore,

involves all four types of investors. Moreover, as is widely observed in practice, investors may change their

investment preferences and decisions from time to time for reasons exogenous to the hedge fund industry

(e.g., risk aversion shocks due to liquidity, human capital, or life-cycle reasons), creating inflows and

outflows. To model this property, we follow Duffie, Gârleanu, and Pedersen (2005, 2007) and assume that

the "h" or "l" intrinsic preference of investors is a Markov chain, switching from low to high with intensity

𝜆 and from high to low with intensity 𝜆 at any time.

Although the involvement of all four types of investors and Markov intrinsic preferences make the

search-and-bargaining process of the existing fund more complicated, we can nonetheless assess fund

incentives by focusing on steady-state conditions under which the fund can raise capital from a stable

composition of investor types for investments. To solve for the steady-state conditions, we first model the

economic reasons for the investor mass to change over time:

⎩⎪⎨

⎪⎧ �̇� = 𝜌 𝜇 + 𝜆 𝜇 − 𝜆 𝜇

�̇� = −𝜇 + 𝜆 𝜇 + 𝜌 𝜇 𝜉(𝛿)

�̇� = −𝜌 𝜇 + 𝜆 𝜇 − 𝜆 𝜇

�̇� = (1 − 𝜆 )𝜇 − 𝜆 𝜇 + 𝜆 𝜇 − 𝜌 𝜇 𝜉(𝛿).

(IA1)

The first equation states that the mass of ℎ𝑜-type investors at any moment during the search process may

change (with a rate of �̇� ) for three economic reasons: 1) when the fund is matched with ℎ𝑛-investors,

which occurs at rate 𝜌 𝜇 , because these noninvestors immediately become ℎ𝑜-investors due to their

higher preference over the expected benchmark return 𝜙 , 2) the transition of investors from 𝑙𝑜- to ℎ𝑜-

type due to Markov intrinsic preference transitions, which occur at rate 𝜆 𝜇 , and 3) the reverse-transition

from ℎ𝑜- to 𝑙𝑜-type, which occurs at rate 𝜆 𝜇 .

Similarly, the second equation says that the mass of 𝑙𝑜-investors may change because (1) these

investors want to quit the fund due to their low intrinsic preference, so the mass changes at rate −𝜇 , (2)

some ℎ𝑜-investors switch their type to 𝑙𝑜 at rate 𝜆 𝜇 , or (3) some 𝑙𝑛-investors are matched and persuaded

to invest—and become 𝑙𝑜-investors at rate 𝜌 𝜇 𝜉(𝛿). The third equation observes that the mass of ℎ𝑛-


investors changes when they are matched with the fund (in which case they become ℎ𝑜-type fund investors),

when some 𝑙𝑛-investors switch their type to become ℎ𝑛-investors, and when the opposite transition

happens. The last equation illustrates that the mass of 𝑙𝑛-investors increases when 𝑙𝑜-investors withdraw

from the fund and become noninvestors (the withdrawal rate is (1 − 𝜆 )𝜇 because a fraction, 𝜆 , of 𝑙𝑜-

investors will switch to high investors rather than withdrawing), decreases when some of them switch to

ℎ𝑛-investors, increases when the reverse switch occurs, and decreases when they are matched with and

persuaded by the fund to become 𝑙𝑜-type fund investors.2

The steady-state of the existing fund is quantified by the conditions �̇� = 0 for each 𝜎 ∈

{ℎ𝑜, 𝑙𝑜, ℎ𝑛, 𝑙𝑛}.3 Hence, we have four equations, in which the economic reasons for changing investor mass

for each type should exactly cancel out in the steady-state. Adding to these questions the condition 𝜇 +

𝜇 + 𝜇 + 𝜇 = 1, we solve for the steady-state investor mass as follows:

⎩⎪⎪⎨

⎪⎪⎧ 𝜇 =

( ( ))

( )[ ( ) ( ) ( )]

𝜇 =( ( ))

( )[ ( ) ( ) ( )]

𝜇 =( )

( )[ ( ) ( ) ( )]

𝜇 =( )

( )[ ( ) ( ) ( )] .

(IA2)

It is convenient to denote 𝐼 = 𝜇 + 𝜇 as the total mass of investors in the fund. Similarly, 𝐼 =

𝜇 + 𝜇 = 1 − 𝐼 denotes the total mass of noninvestors in the fund. From the above solution, we have

𝐼 =( )( )

( )[ ( ) ( ) ( )]

𝐼 = 1 − 𝐼 =( ) ( ) ( )

( )[ ( ) ( ) ( )]. (IA3)

It is easy to see that 𝐼 strictly decreases in 𝜉(𝛿): < 0 as long as we have nonzero parameters. Hence,

𝐼 strictly increases in 𝜉(𝛿): =( )

> 0. This property is intuitive: more persuasion will allow the

fund to attract a larger investor base and thus more capital. This property will be useful later when we

examine managers’ incentives.

Based on the above steady-state conditions on investor clientele, the capital raised by the benchmark

fund is 𝑧(𝑡)𝐼 . Condition (3) of new funds becomes the following for the benchmark fund:

𝑧(𝑡)𝐼 ≥ 𝑞 (𝑡). (IA4)

2 Note that the set of conditions (IA1) does not include the influence of the new fund, because the existing fund moves first to lock in the capital for investment. 3 Similar to the bargaining process of Rubinstein and Wolinsky (1985), the stationary conditions can be regarded as a “market equilibrium” in the sense that they describe the strategies of market participants with both semi-stationarity (i.e., independent of history) and sequential rationality (i.e., the strategies are optimal after possible histories).


Since Lemma IA2 focuses on the benchmark fund, we omit the superscript “E” hereafter when there is no

risk of confusion. Similar to the new manager, the manager of the benchmark fund (hereafter, the existing

or old manager) decides the optimal level of search cost and extra performance as follows:

max( ),

𝑈(𝑠(𝑡), δ) (IA5)

s. t. 𝐸(𝑟 ) = 𝜙 − 𝑐(𝑞 ) − 𝑠(𝑡) = 0 and 𝑧(𝑡)𝐼 ≥ 𝑞(𝑡),

where 𝑈(𝑠(𝑡), δ) = 𝑔(𝑞 ) − 𝐿(δ) = 𝑓 × 𝑞 − 𝐿 × δ, and 𝑠(𝑡) =( )

( )=

× ( ) is the fund-size scaled

search cost.

Our objective is to show that, under reasonable market conditions, a steady-state exists in which the

existing fund manager does not have the incentive to generate additional alpha, even when upon similar

conditions the new manager may choose to generate alpha according to Proposition 1. We demonstrate this

property in two steps. First, we solve for the optimal search cost that the existing fund wants to pay without

generating any alpha. We denote the steady-state of this equilibrium as the search-only-equilibrium. The

optimal scaled search cost, matching intensity, fund size, and alpha are referred to as 𝑠∗, 𝜌∗, 𝑞∗, and δ∗ for

this equilibrium (of course, δ∗ = 0). Second, we establish the conditions under which any additional effort,

𝑑𝛿, as a deviation to the search-only-equilibrium, leads to disutility for the manager as a first-order effect.

Hence, the manager has no incentive to deviate. Note that these conditions establish the existence of the

search-only-equilibrium for the existing fund but do not rule out possible multiple equilibria. Since our

numerical calibration (presented shortly) suggests that these conditions are likely to hold in practice, we

rely on this equilibrium as a benchmark to understand the performance incentives of new fund managers.

The optimal search cost in the search-only-equilibrium can be derived as follows. Without alpha,

existing managers cannot persuade low-noninvestors to invest in their fund; mathematically, 𝜉(𝛿) = 0. In

this case, 𝐼 =( )

( )[ ( )] from solution (IA3). Next, similar to the case of Proposition 1, it is

optimal for the existing fund to pay a search cost to attract exactly the amount of capital optimal for

operation: that is, 𝑧(𝑡)𝐼 = 𝑞(𝑡), from which we can derive 𝜌∗ =( ) ( ) ( )

( ) ( ) ( ) ( ) and 𝑠∗(𝑡) =

∗

( )=

( ) ( )

[ ( ) ( ) ( ) ( )].

Plugging this steady-state search cost into the BG condition 𝜙 − b × 𝑞 − 𝑓 − 𝑠(𝑡) = 0, we solve for

the fund size as 𝑞∗(𝑡) = +( ) ( )

− −( ) ( )

+( )

.4 Furthermore, the

4 Note that the BG condition implies a quadratic function, 𝑞∗(𝑡), and thus has two potential solutions. But only this one has proper

economic meaning. The other solution, 𝑞∗(𝑡) = +( ) ( )

+ −( ) ( )

+( )

, implies a negative

search cost and ends up with a larger fund size than , the fund size of BG without the search.


condition for fund size to be positive (i.e., 𝑞∗(𝑡) > 0) implies that +( ) ( )

>

−( ) ( )

+( )

, which is equivalent to 𝜙 > 𝑓 +( ) ( )

( ) ( ). That is, the

performance of the existing fund should cover both the management fees and a minimum threshold of

search cost to allow for positive fund size in the BG equilibrium.

We next examine the conditions under which the manager has no incentive to deviate from the search-

only equilibrium by making a small effort, dδ. From (IA5), this occurs when the learning cost of dδ exceeds

its marginal benefit, or 𝑓 × 𝑞′ ≤ 𝐿 . Since the BG condition has the same functional form for both new

and existing managers, the marginal BG condition = − derived in Proposition 1 also applies to the

benchmark fund, from which we can rewrite the no-effort condition as − ≤ .

To solve the no-effort condition, we need to calculate . Rewriting the search cost as 𝑠(𝑡) =∗

( )=

∗

( )( ), we derive

( )=

( )=

∗

( )( )× × 𝜉 . Since 𝐼 =

( )( )

( )[ ( ) ( ) ( )], we have = −

( )( )

( )[ ( ) ( ) ( )]× (𝜌 + 𝜆 )𝜌 = −𝐼 ×

( )( )

( )( ). This allows us to write

( ) as

( )= −

( )( )

( ) ( )( ) ( )=

−( )( )

( ) ( )( )

( ). Plugging

( ) back into the no effort condition, we have

( )( )

( ) ( )( )

( )≤ . Since

( ) is decreasing in 𝐼 , the no-effort condition holds when

𝐼 ≥ 1/ 1 +( ) ( )( )

( )( )≡ 𝐼 , where 𝐼 is a threshold of 𝐼 . Because 𝐼 < 1, there

always exist values of 𝐼 for which the no-effort condition holds—this condition holds in the numerical

example that we will discuss shortly. This result is intuitive: since the pool of existing investors incurs no

additional search cost, when its mass gets big enough (i.e., exceeding a certain threshold), the extensive

margin becomes high enough for the existing fund to focus on high-type investors. In this case, the fund

will have little incentive to resort to additional performance to persuade low-type investors. This proves

Lemma IA2.5

D. Proof of Proposition 2

The problem to be solved by the new affiliated fund manager is similar to that of the new stand-alone

manager, except that the density of noninvestors for the former is 𝜇 ′ (𝑡) = 𝜇 (𝑡) + 𝛾 𝜇 (𝑡), where

5 Note that the incentives of existing funds will not be influenced by the potential competition introduced by high-alpha new funds, because the high-alpha inception will have existing investors and lower search costs in the next period and become an existing fund (as opposed to motivating the previous existing fund to deliver alpha).


𝛾 > 0 is the networking parameter. Following the spirit of Proposition 1, we can derive the new optimal

search cost as 𝑠 ′(𝑡) =( )[ ′ ( ) ( ) ( )]

for affiliated inceptions. Compared to the search cost of

stand-alone inceptions, 𝑠 (𝑡) =( )[ ( ) ( ) ( )]

, we see that 𝑠 ′(𝑡) < 𝑠 (𝑡), confirming that the

affiliated inception pays a lower search cost than a stand-alone inception. This proves property i.

Following the proof of Proposition 1, we derive new FOC as 𝐿 = ∙( )

( ) ′ ( ) ( ) ( ), which

gives the optimal level of performance as 𝛿′∗

= 𝑀𝑎𝑥{0, 𝑓 (𝐿 𝑏𝜇 (𝑡)𝜉 𝑎𝑧 ) −′

( )

( )}. Compared to

the optimal alpha of stand-alone funds, 𝛿∗

= 𝑀𝑎𝑥{0, 𝑓 (𝐿 𝑏𝜇 (𝑡)𝜉 𝑎𝑧 ) −( )

( )}, it is

straightforward to see that 𝛿 ′∗

≤ 𝛿∗ in general and that the former is strictly smaller than the latter when

the latter is nonzero. This proves property ii. Property iii follows from the closed-form solution of 𝛿 ′∗.

The optimal fund size can be solved as 𝑞 ′∗(𝑡) = −( )

for nonzero 𝛿 ′∗.

Finally, as discussed in proposition i and Lemma IA1, convex learning costs and the retention of a

fraction of performance to either relax the BG condition by offsetting operational costs or as incentive fees

will not change the main conclusions of this proposition.

E. Lemma IA3 and its Proof:

LEMMA IA3: Search frictions amplify diseconomies of scale for the existing hedge fund, and this effect

increases with fund size. Due to this property, when an existing fund encounters excess demand, it is optimal

for the fund to launch an affiliated new clone fund to accept additional capital. Clone funds launched in

this way have little incentive to deliver extra performance.

Proof: First, we note that < 0; hence, the search cost paid by the existing fund has a diminishing benefit

as fund size increases. The intuition of this effect is as follows. By growing bigger or spending more on the

search cost, the fund effectively transfers more high-type noninvestors into high-type investors. But doing

so also reduces the fraction of remaining high-type noninvestors in the market (i.e., from equation IA2,

= , which decreases in 𝜌) for future search, which lowers the marginal benefit of additional

searching. Note that the diminishing benefit of search costs is equivalent to an increasing and convex

relation between search costs and fund size (we can show that ∗

> 0 and ∗

> 0).


We now prove that search frictions amplify the diseconomies of scale for the existing hedge fund and

that this effect increases in fund size. To do so, we rewrite the search-enhanced BG condition as 𝜙 − (b +∗

) × 𝑞 − 𝑓 = 0 or 𝜙 − b × 𝑞 − 𝑓 = 0 (where b = b +∗

parameterizes the search-amplified

diseconomies of scale) and compare it to the BG condition of mutual funds without search (i.e., 𝜙 −

b × 𝑞 − 𝑓 = 0). Since ∗

> 0, we have b > 𝑏: search frictions amplify the diseconomies of scale for the

existing hedge fund. Moreover, ( )

=∗

> 0, so the diseconomies-of-scale amplification effect—the

difference between the hedge fund diseconomies of scale and those implied by the BG model without search

frictions—also increases in fund size due to the diminishing benefit of search costs.

The overall amplification effect of search frictions on diseconomies of scale suggests that hedge funds

are willing to maintain smaller fund sizes than they would without these frictions. Moreover, this effect

provides a rationale for the widely observed inceptions of clone funds in the hedge fund industry. To see

the result, we simplify the BG condition as 𝜙 − 𝑓 = 𝑤(𝑞 ), where 𝑤(𝑞 ) = b × 𝑞 + 𝑠∗(𝑞 ) is a convex

function of fund size 𝑞 (because ∗

> 0,∗

> 0). Consider the case in which an excess amount of capital,

Δ𝑞 , suddenly becomes available to an existing fund with a steady-state fund size 𝑞∗ and corresponding

search cost 𝑠∗. In our model, this type of demand shock can occur due to a temporary shift in investor

preferences or to the Markov probability of preference change. Jensen’s Inequality implies that 𝜙 − 𝑓 =

𝑤(𝑞∗ + Δ𝑞 ) > 𝑤(𝑞∗) + 𝑤(Δ𝑞 ). It is optimal to use a clone fund managed by a new manager to absorb

the excess capital.

Although amplified diseconomies of scale may be only one attribute to the inception of clone funds,

its prediction for inception performance is unambiguous and can be tested empirically. Specifically, since

the clone fund adopts the same investment strategy as the existing fund by design, it has no incentive to

deliver performance above and beyond its preceding fund.

Before we provide a numerical example, we further discuss the assumptions of the model. First, similar

to the bargaining process of Rubinstein and Wolineky (1985), the stationary conditions can be regarded as

a “market equilibrium” in the sense that they describe the strategies of market participants with both semi-

stationarity (i.e., independent of history) and sequential rationality (i.e., the strategies are optimal after

possible histories). Second, we have assumed that fund managers have linear bargaining power built on

performance. As noted in footnote 10 of the main article, this reduced-form assumption is consistent with

both the BG equilibrium concept and the take-it-or-leave-it Nash bargaining models widely used in the

search literature. To illustrate the latter point, we sketch a general framework that can generate results

consistent with both the assumption of having two different types of investors and the assumption of

increasing bargaining power based on superior performance.


Assume that there exists a continuum of investors with a uniformly distributed discount factor 𝛽

between zero and one. In each period, investors can decide to invest in a fund, which delivers an expected

return of 𝜙 + 𝛿 in the next period that can generate a von Neumann and Morgenstern utility of 𝛽(𝜙 +

𝛿) for investors. Alternatively, investors can reject the fund and immediately receive a non-negative utility

from an outside option. Two general results arise in the Nash equilibrium following the reservation-value

rule. First, there exists a cutoff point, 𝛽, above which investors accept the fund. Second, an increase in δ

increases the mass of investors investing in the fund, which gives rise to our bargaining power assumption.

Although our stylized model focuses on reduced-form assumptions for simplicity, these assumptions can

be reasonably derived in a more general search model.

F. A Numerical Example of the Manager-Search Model

In this section, we provide a numerical example of the manager-search model. We select parameter

values such that the model-implied fund size and the performance of both the existing benchmark fund and

the stand-alone new fund can be matched with, respectively, the average fund size and the performance of

existing funds and new funds observed in our data. This calibration exercise allows us to visualize the

relationship between inception incentives in the intensive margin and variation in investor demand in the

extensive margin, as Figure 1 demonstrates.

We first describe the size and performance of existing and new funds to be matched in our model. The

size of existing individual hedge funds is measured by the average AUM of existing funds across our

merged TASS/HFR/BarclayHedge database, which is $208M. This value is close to what has been reported

in existing studies.6 From Table IV, existing hedge funds deliver an after-fee risk-adjusted performance of

2.72% per year (0.224% per month) on average. Next, a typical new fund in our sample raises about $44.6M

in capital at its inception. Finally, a typical stand-alone new fund delivers an additional risk-adjusted

performance of 0.458% per month in Table V (also after fees), which outperforms existing funds by 0.234%

per month or 2.84% per year, whereas a typical family-affiliated nonclone inception delivers 0.258%

monthly alpha in our sample, which outperforms existing funds by 0.034% per month or 0.409% per year.

These performance differences are what fund-specific alpha in Propositions 1 and 2 aims to capture.

Table IAI summarizes the parameters in our calibration exercise. Observed values from the data are

tabulated in Panel A. Panel B summarizes the parameters selected to match fund characteristics for existing

and stand-alone inceptions, and Panel C contains the calibrated parameters of the steady-state of existing

funds. Panel D reports the parameter selected to match the extra performance of family-affiliated nonclone

6 For instance, the average AUM is about $213M in Jaganathan, Malakhov, and Novikov (2010) and calibrated as $206M in Glode and Green (2011).


inceptions. Below we describe our calibration process based on Proposition 1 (the optimal alpha of stand-

alone new funds), Lemma IA2 (the steady-state condition of the benchmark fund), and Proposition 2 (the

optimal alpha of family-affiliated new funds), as well as the economic interpretation of the parameter values

used in this exercise.

We first select a few parameters to match the size and performance of the existing funds (Panel B of

Table IAI). The annual management fee and the endowed strategy performance of all funds are set to 2%

and 4.72%, respectively, to match the 2.72% after-fee performance of existing funds. These two parameters,

together with the parameter on diseconomies of scale, 𝑏, jointly imply a fund size of $226 million in a

straightforward BG equilibrium without search frictions. The difference between this number and actual

fund size comes from search frictions.

In particular, the existing fund manager needs to search due to fund outflows. Fund flows are

determined by the Markov transition probabilities of low-preference investors switching to high-preference

and vice versa, both of which are assumed to be 0.22 over a year (the actual fund outflow will depend on

the steady-state composition of high- versus low-type investors, as we discuss below). The capital available

to an average individual fund, 𝑧(𝑡), is assumed to be $1.6 billion with a matching intensity, 𝜌, of 0.1. These

two parameters are selected to allow a typical fund to raise up to $160 million of capital in the search step;

this total amount is between the observed AUM of existing funds and that of new funds in order to match

the size of these two types of funds. Based on these parameters, the search efficiency (i.e., parameter 𝑎 )

is further selected to imply a reasonable total search cost of about $1 million for a manager to meet with all

potential investors. The benchmark fund does not need to spend that much to search for all possible

capital—it only needs to find enough capital from noninvestors to offset the outflows of existing investors.

To shed more light on outflows, in Panel C we report the distribution (mass) of each type of investor

in the steady-state of the existing benchmark fund can be calculated from the above set of parameters.

According to Lemma IA2 (equations IA2 and IA3), the benchmark fund attracts 22.5% of potential

investors to invest in the fund, among which about 82% are high-preference (i.e., 18.4%/22.5% = 82%)

and the rest are low-preference. This investor composition implies that 18% (i.e., 82% × 0.22 = 20%) of

high-type investors will convert to low-type investors, whereas 4% (i.e., 18% × 0.22 = 4%) of low-type

investors will have the opposite conversion. The difference between these two conversions implies a net

outflow of about 14%, which requires that the manager of the existing benchmark fund search for new

investors. In this steady-state, an existing fund raises about $208M in capital according to Lemma IA2 (our

baseline case), which matches data-implied fund size.

We next discuss the properties of the new fund. Note that because the existing benchmark fund has

already attracted a fraction of investors, the capital-raising process of the new fund is conditioned on the

investors not retained by the existing fund. In other words, although 𝜇 and 𝜇 are endogenous to the


existing fund (i.e., a higher search cost paid by the fund will leave less high-type investors in the market),

they are the exogenous market conditions for the new fund. Hence, we need to apply the 𝜇 and 𝜇

calibrated from Lemma IA2 to new funds according to Proposition 1.

Based on the steady-state distribution of investors as reported in Panel C, we further select the

bargaining efficiency and learning cost parameter values to match the size and performance of new hedge

funds according to Proposition 1. These parameter values are also reported in Panel B. The calibrated

bargaining efficiency parameter, 𝜉 , suggests that for every 1% additional performance, a fund manager

can persuade 2.9% of otherwise suspicious investors to invest in the fund. Meanwhile, the implied learning

cost is about $0.12 million for every 1% of risk-adjusted return delivered. These additional parameters

allow the baseline stand-alone new funds to generate an average alpha of 2.84% above existing funds with

an inception AUM of about $44 million according to Proposition 1.

Overall, the parameters reported in Panel B closely match the observed size and performance of

existing and new stand-alone hedge funds. We therefore treat this parameter set as a baseline case

associated with the average incentives of new stand-alone funds. This numerical example allows us to

demonstrate how deviations from the baseline case—in terms of both investor demand and family

structure—alter the incentives of new hedge fund managers.

Our model first predicts that the incentives of new fund managers differ according to the extensive

margin of the search step. To demonstrate this relationship, we measure fund incentives as model-implied

optimal performance that the new fund manager is willing to deliver, and we focus on exogenous variations

in the extensive margin introduced by z(t), that is, when this variable deviates from the above baseline case

value, denoted by z. The investor demand index, z(t)/z, then captures the exogenous variation—again from

the new manager’s perspective—in the total amount of capital carried by investors into the hedge fund

industry.

Figure 1 illustrates how investor demand affects performance by plotting the optimal performance, δ∗,

for stand-alone inceptions, family-affiliated nonclone inceptions, and family-affiliated clone inceptions

according to, respectively, Proposition 1, Proposition 2, and Lemma IA3. For stand-alone inceptions, a 12%

increase in investor demand (i.e., 𝑧(𝑡)/𝑧 = 1.12) suffices to wipe out the alpha incentive of the new fund

managers, whereas a 6.7% decrease in investor demand will incentivize the new fund to deliver fund alpha

that matches the performance of cold stand-alone inceptions as reported in Table VI: these inceptions

deliver a risk-adjusted return of 7.44% per year (0.60% per month), implying an additional alpha of 4.72%

per year above the 2.72% performance of existing funds. Hence, the performance of cold and hot hedge

fund inceptions observed in our empirical analyses can be generated from very reasonable ranges of demand

changes.


Our model next predicts the existence and influence of family structure. In this example, the rationale

of the family structure can be seen from the steady-state distribution of investors in Panel C. There, the high

investor-to-low investor ratio retained by the existing fund (i.e., 18.4% to 4.1%) is much higher than that

of noninvestors still available in the market (i.e., 45.9% to 31.6%). This difference arises because the

existing fund has already paid search costs to attract high-type noninvestors to invest in the fund. Because

the affiliated new manager benefits from the pool of investors of the existing fund through the family

structure, she enjoys a much higher extensive margin, which saves search costs but lowers the incentive of

the manager to deliver performance. This networking effect gives rise to family-affiliated nonclone

inceptions.

To quantify the effect, we first add the networking effect into the baseline case, and then select the

value of the networking parameter, 𝛾 , to match the 0.409% alpha spread between nonclone inceptions and

existing funds according to Proposition 2. Based on the calibrated 𝛾 value (reported in Panel D of Table

IAI), Figure 1 then plots the relationship between investor demand, z(t)/z, and the optimal fund-specific

alpha of family-affiliated nonclone inceptions. The figure shows that, in this numerical example, a 2%

increase in investor demand is enough to wipe out the alpha incentive of the affiliated new manager. In

contrast, a 9% decrease in investor demand will incentivize the new nonclone fund to generate the

performance of cold nonclone inceptions as reported in Table VI: these inceptions deliver a 5.7% risk-

adjusted return per year (i.e., 0.463% per month), which is 2.98% above what existing funds generate.

Figure 1 finally illustrates the incentives of clone funds as described in Lemma IA3. Since a clone fund is

launched to absorb the large demand shock of the existing fund, the clone fund will not have the incentive

to deliver performance above that of the existing fund. Hence, its optimal fund-specific alpha is plotted as

zero.


Table IAI Parameter Values for the Baseline Case of the New and Existing Funds

This table reports the parameter values used in our numerical example. The parameters are selected to match the observed characteristics of the hedge fund industry, including the size and risk-adjusted performance of new and existing funds. In Panel A, the column “Data” tabulates the observed values of the fund variables to be matched, whereas the next column demonstrates the overall fit of the model. Panel B tabulates the value of the parameters used in Proposition 1 and Lemma 2 to achieve this fit. Panel C reports the steady-state distributions of investor types, which are calibrated from the existing funds according to equations (IA2) and (IA3) in Lemma 2 and then applied to the new fund following Proposition 1. The values of exogenous variables used to fit the size and performance of the new fund are also tabulated in Panel B. Panel D reports the parameter for the networking effect of family-affiliated inceptions.

Symbol Data Model

Panel A. Fund characteristics to be matched

The benchmark fund size (in $mn) 𝑞 208.0 208.0 After-fee benchmark fund risk-adjusted return (annualized) 𝜑 − 𝑓 2.72% 2.72% New fund size (in $mn) 𝑞 44.6 44.0 Additional risk-adjusted return generated by stand-alone inceptions (Proposition 1)

δ (stand-alone)

2.84% 2.84%

Additional risk-adjusted return generated by nonclone inceptions (Proposition 2)

δ (nonclone) 0.409

% 0.409

% Panel B. Parameter values selected to match fund characteristics of stand-alone and existing funds

Expected annual benchmark performance φ 4.72% Management fee f 2.00%

Diseconomy of operation scale b 0.00012

Search efficiency a 0.10 Amount of capital carried by investors (available to the fund, $mn) z 1600 Matching intensity 𝜌 0.10 Bargaining efficiency 𝜉 2.9

Learning Cost 𝐿 11.74 Transition probability from low to high 𝜆 0.22 Transition probability from high to low 𝜆 0.22

Panel C. Steady-state variables (calibrated from the benchmark fund, applied to the new fund)

The mass of noninvestors with high-preference 𝜇 45.9%

The mass of noninvestors with low-preference 𝜇 31.6%

The mass of existing investors with high-preference 𝜇 18.4%

The mass of existing investors with low-preference 𝜇 4.1%

The mass of noninvestors 𝐼 77.5% The mass of existing investors 𝐼 22.5%

Panel D. Parameter values selected to match the performance of family-affiliated nonclone inceptions

The parameter of networking effect 𝛾 12.1%


II. An Investor Search-Enhanced Berk and Green (2004) Model

In the main text, our extension of the Berk and Green (2004) model features new hedge fund managers

searching for sophisticated investors. In practice, search frictions are likely to occur in both directions (i.e.,

both investors and managers may need to engage in costly search). Here we provide an extension of Berk

and Green (2004) based on the investor search framework of Hortaçsu and Syverson (2004). We then

discuss the implications of this model in terms of hedge fund inceptions.

Our investor-search framework implies very high search frictions for investors seeking new funds,

suggesting that new funds could benefit tremendously from spending resources on reaching out to investors.

This feature suggests that the two search frameworks (investor search and manager search) complement

each other in describing the hedge fund industry and that manager-search could play an essential role in

explaining the economics of hedge fund startups.

A. A Simple BG Framework of Investor-search

A.1. The Setup of the Model: Fund Managers

We start with the description of 𝑁 funds within a particular benchmark style. Each fund is endowed

with a benchmark risk-adjusted return, 𝑅 , with an expected value of 𝜙 in period 𝑡. Note that the

benchmark return is the same for all funds in this style. To simplify the notation, we omit the index 𝑡 when

there is no risk of confusion. We further assume that fund 𝑗 will distribute a cost-adjusted return 𝑟 = 𝑅 −

(𝑓 + 𝑏 × 𝑄 ), where 𝑄 refers to fund size, and 𝑏 and 𝑓 refer to operational cost (with diseconomies of

scale) and management fees, following Berk and van Binsbergen (2015, 2017).

It is convenient to present fund size in terms of the market share. We denote the total amount of capital

carried by all investors by 𝑍 and use 𝑞 to denote the fractional market share occupied by fund 𝑗, that is,

𝑄 = 𝑞 × 𝑍.

A major difference between the original Berk and Green (2004) model and the investor-search model

is that fund size is determined by the search process of investors in the latter case, as we will show below.

We assume that 𝜙 − 𝑏 × 𝑄 > 0, such that each fund manager is able to deliver a nonnegative return to

investors based on the endowed benchmark performance.

𝐸 𝑟 = 𝜙 − (𝑓 + 𝑏 𝑞 𝑍) ≥ 0. (IA6)


Next, as in the model in our main text, we allow the fund to acquire additional fund-specific

performance through costly investment in learning. This allows the fund to deliver additional performance

to investors (as compensation for search costs in equilibrium). In particular, the fund can choose to pay a

private learning cost, 𝐿 = 𝐿 𝑍𝛿 /2, which generates expected extra performance 𝛿 . Here, the cost function

is assumed to be quadratic. We also assume that the cost is the same for all managers and is proportional to

the total size of the capital market, which captures the idea of Gârleanu and Pedersen (2018) that a larger

fund sector makes the market more efficient (and hence increases the information cost).

Once the extra fund-level performance is generated, the fund manager delivers 𝛿 to investors as

compensation to their search costs. Since the fund generates an expected return of 𝜙 + 𝛿 for the period

and delivers 𝛿 to investors in addition to what it previously delivers, condition (1) holds: since investors

get 𝐸(𝑟 ) = 𝜙 + 𝛿 − (𝑓 + 𝑏 𝑞 𝑍) ≥ 𝛿 , we still get 𝜙 − (𝑓 + 𝑏 𝑞 𝑍) ≥ 0.

We assume that the fund manager can derive a utility gain of 𝑓 × 𝑄 by managing a fund with size 𝑄 .

Hence, the fund manager maximizes the learning-cost-adjusted utility as follows:

𝑚𝑎𝑥,

𝑈 𝑓 , 𝛿 = 𝑓 𝑄 − 𝐿(𝛿) = 𝑓 𝑞 𝑍 − 𝐿 𝑍𝛿 (IA7)

s.t. (IA6)

The above managerial problem assumes that funds have heterogenous operational cost (𝑏 ) but a

homogeneous learning cost coefficient (𝐿 ). The heterogeneity in operational cost, as we will see below,

allows low-cost funds to deliver higher optimal performance in equilibrium.7

A.2. Investors and Searching Costs

We now introduce the setup of investors searching for funds. We assume that the economy features a

continuum of investors with heterogeneous searching costs. Before we get into the details, some intuition

about the investor-search framework is helpful.

Suppose there are N funds ranked by their performance 𝛿 < 𝛿 < ⋯ < 𝛿 . If we denote 𝑐 as the

lowest possible search cost of any investor to invest in fund 𝑗, then the search equilibrium creates a set of

7 We can, of course, also allow for heterogeneous learning costs, but the main feature of the equilibrium will not change as long as some funds have a cost advantage compared to others. Hence, for tractability, we focus on heterogeneous operational costs in our analysis. Out of the three fund policies, 𝜙 , 𝑓 , and 𝛿 , there are two degrees of freedom due to the constraint of equation (IA6) and multiple ways of modeling fund policies. Our current setting implies that new funds will adopt the same level of fees as the existing fund if their operational costs are similar. But they will pay different private costs to acquire more skill. Allowing funds to select interior fees will not change the key features of the model.


cutoff search costs such that 𝑐 > 𝑐 > ⋯ > 𝑐 , following the optimal search rule of investors and the

optimal performance delivered by funds. Intuitively, investors with high search costs are stuck with low-

performing funds, and investors with low search costs are able to find better-performing funds.

With this reverse match between search costs and fund performance in mind, we can use the general

distribution function 𝐺(𝑐) to describe the cumulative mass of investors whose search costs are lower than

𝑐. The corresponding distribution intensity is denoted as 𝑔(𝑐) (i.e., 𝑔(𝑐) = 𝑑𝐺(𝑐)/𝑑𝑐). We restrict each

investor to invest in only one fund.

We further assume that investors derive the utility of 𝛿 from fund 𝑗. This linear utility simplifies the

mathematics, such that spending on search costs can be directly compared to gains from expected fund

performance.

We follow Hortaçsu and Syverson (2004) and focus on sequential search, where investor 𝑟 with search

costs 𝑐 finds and invests in funds as follows. For each search, the investor is randomly matched with a

fund. If the fund is new (i.e., not matched before), the investor pays a learning/searching cost 𝑐 to learn

about its type (i.e., 𝛿 ). If the investor has been matched with the fund before, then the investor does not

need to pay additional search costs. The investor then decides whether to invest. If the investor invests, the

search process ends. If not, investors search for another random fund. At any time, the investor can revisit

the matched fund(s).

Hortaçsu and Syverson (2004) describe the assumptions leading to the existence of an optimal search

rule under the above sequential search. We assume that all of these conditions hold in our economy and

thus focus on the optimal rule, which gives the cutoff points 𝑐 for fund 𝑗 as follows:

𝑐 = 𝜌 (𝛿 − 𝛿 ), (IA8)

where 𝜌 is the sampling probability for fund 𝑘 to be randomly matched with an investor in any random

search (public information). In equation (IA8), ∑ 𝜌 (𝛿 − 𝛿 ) is the expected benefit for an investor

already matched with fund 𝑗 to conduct another search. When the benefit equals the search cost of the

investor, the investor will stop searching. Any investors with 𝑐 ≥ 𝑐 matched with fund 𝑗 will end up

investing in the fund. Hence, 𝑐 denotes the lowest possible search cost of any investor considering the

investment in fund 𝑗. Note that equation (IA8) also implies that 𝑐 = 0.

From the cutoff points 𝑐 , the market share of each fund can be calculated. For instance, 𝜌 (1 − 𝐺(𝑐 ))

determines the market share of the first fund with the lowest performance, 𝛿 . This result arises because


only 𝑐 ≥ 𝑐 investors (whose aggregate mass is 1 − 𝐺(𝑐 )) matched with this fund in their first random

search will invest in the fund. If these investors are initially matched with a better fund, they will stay with

that fund. In other words, these investors do not expect to benefit from a second search based on equation

(IA8). In contrast, if 𝑐 < 𝑐 investors are initially matched with this 𝛿 fund in their first search, they will

continue to search util they meet a new (and better) fund.8

To determine the market share of the second-worst fund, 𝛿 , we need to consider not only initially

matched investors but also investors who had been initially matched with 𝛿 but can find this fund in their

future search. This intuition allows Hortaçsu and Syverson (2004) to derive the closed-form solution for

market shares, which also applies to our case. The market share of fund 𝑗 is

𝑞 = 𝜌 1 +𝜌 𝐺(𝑐 )

1 − 𝜌+

𝜌 𝐺(𝑐 )

(1 − 𝜌 )(1 − 𝜌 − 𝜌 )+

𝜌 𝐺(𝑐 )

(1 − 𝜌 − ⋯ − 𝜌 )(1 − 𝜌 − ⋯ − 𝜌 )

−𝐺 𝑐

1 − 𝜌 − ⋯ − 𝜌. (IA9)

A.3. Optimal Fund Performance

We now examine the conditions under which a search equilibrium arises under optimal fund policies

and optimal investor search rules. Below, we first derive the manager’s optimal solution in Proposition IA1.

We then quantify the sufficient condition for performance heterogeneity to arise in equilibrium based on

cost advantages.

PROPOSITION IA1: When fund manager 𝑗 solves problem (IA7) conditioning on investors’ optimal search rule (IA8) and the resulting market share condition (IA9), the optimal fund policies are 𝑓∗ = 𝜙 −

𝑏 𝑞 𝑍, and

𝛿∗ =1

𝐿𝑓∗ − 𝑏 𝑞 𝑍

𝑑𝑞

𝑑𝛿

8 Note that Hortaçsu and Syverson (2004) assumes that there are no costs when the search ends up with a matched fund. Hence, 𝑐 < 𝑐 investors initially matched with this fund will search and find a better fund.


=1

𝐿𝜙 − 2𝑏 𝑞 𝑍

𝜌 𝜌 𝑔(𝑐 )

1 − 𝜌+


(1 − 𝜌 )(1 − 𝜌 − 𝜌 )

+𝜌 𝜌 𝑔(𝑐 )

(1 − 𝜌 − ⋯ − 𝜌 )(1 − 𝜌 − ⋯ − 𝜌 )

+𝜌 ∑ 𝜌

1 − 𝜌 − ⋯ − 𝜌𝑔 𝑐 . (IA10)

Proof: The manager will use fees to absorb all economic rents, such that 𝜙 − 𝑓∗ + 𝑏 𝑞 𝑍 = 0. Hence,

𝑓∗ = 𝜙 − 𝑏 𝑞 𝑍. Since all funds have the same benchmark performance 𝜙, the economic rents depend on

fund size and operational costs.

In this case, 𝑈 𝑓∗, 𝛿 = (𝜙 − 𝑏 𝑞 𝑍)𝑞 𝑍 − 𝐿 𝑍𝛿 /2. The FOC of 𝑈 𝑓∗, 𝛿 leads to 𝛿∗ =

𝜙 − 2𝑏 𝑞 𝑍 . To derive , we plug in equation (IA9) and notice that ( )

= 𝑔(𝑐 ) . To get

, we have from equation (IA8) that

𝑑𝑐

𝑑𝛿=

⎩⎪⎨

⎪⎧

0, 𝑖𝑓 𝑟 > 𝑗

− 𝜌 , 𝑖𝑓 𝑟 = 𝑗

𝜌 , 𝑖𝑓 𝑟 < 𝑗

which allows us to solve for as follows

𝑑𝑞

𝑑𝛿=


1 − 𝜌+


(1 − 𝜌 )(1 − 𝜌 − 𝜌 )+


(1 − 𝜌 − ⋯ − 𝜌 )(1 − 𝜌 − ⋯ − 𝜌 )

+𝜌 ∑ 𝜌

1 − 𝜌 − ⋯ − 𝜌𝑔 𝑐 . (IA11)

Plugging equation (IA11) into the FOC, we get equation (IA10).

The intuition of equation (IA10) is most clearly demonstrated in an economy when 𝜌 = 1/𝑁 for all

funds (i.e., when the sampling probability is the same for all funds). In this case, we can rewrite equation

(IA10) as

𝛿∗ = 𝑓 ×𝑑𝑞

𝑑𝛿=

1

𝐿𝜙 − 2𝑏 𝑞 𝑍 × 𝛽 × 𝑔(𝑐 ) , (IA12)


where 𝑓 = 𝜙 − 2𝑏 𝑞 𝑍 and the 𝛽 coefficient in is defined as

𝛽 =

⎩⎨

⎧1

𝑁(𝑁 − 𝑘)(𝑁 − 𝑘 + 1), 𝑘 = 1, … , 𝑗 − 1

𝑁 − 𝑗

𝑁(𝑁 − 𝑗 + 1). 𝑘 = 𝑗

(IA13)

In equation (IA12), 𝑓 denotes the economic rents the fund manager can collect per dollar invested in

the fund after adjustments for learning and operational costs. Proposition IA1, therefore, states that the

higher the economic rents are, the stronger incentive the manager has to generate higher fund-specific

performance.

The next term, = ∑ 𝛽 × 𝑔(𝑐 ), denotes the marginal impact of performance on fund size. There

are two important properties of this marginal benefit. First, both equations (IA11) and (IA13) imply that

> 0, in other words, better performance attracts more capital. Second, 𝛽 is an increasing function of

𝑘. As a result, among all types of investors investing in fund 𝑗, investors with lower search costs play the

most important role. Indeed, investors with the lowest search costs (i.e., at or close to 𝑐 ) are the marginal

investors to the fund.

Jointly, the two properties imply that any marginal performance improvement will attract more

marginal investors, which leads, in turn, increased fund size. Of course, additional performance means

higher learning costs as well. The balance between the benefits and costs allows fund managers to find an

interior optimal solution.

The above discussion suggests that this marginal impact plays a similar role in the search equilibrium

as the flow-performance sensitivity does in the original Berk and Green (2004) model. Both allow the

market to equilibrize through quantity (i.e., investor capital). Of course, in our search friction framework,

all capital is not the same: low-cost investors (i.e., marginal investors) have a higher impact on the

sensitivity between market-share changes (the equivalence of flows in our framework) and performance.

Due to investor heterogeneity, therefore, the flow-performance sensitivity with search frictions remains

positive but is weaker than that of Berk and Green (2004).

A.4. Cost Advantage as the Economic Ground of Performance Heterogeneity

Although Proposition IA1 quantifies the optimal policies of funds, the search equilibrium further

requires heterogeneous performance (i.e., 𝛿∗ < 𝛿∗ < ⋯ < 𝛿∗ ). Economically speaking, the cost-benefit of


fund 𝑗 + 1 should be large enough compared to fund 𝑗 that the former fund optimally delivers higher

performance despite costly learning. Below we quantify the conditions of the cost-benefit that allow

performance heterogeneity to arise in equilibrium.

PROPOSITION IA2: 1) For fund 𝑗 + 1 to optimally deliver better performance than fund 𝑗 in equilibrium (i.e., 𝛿∗ < 𝛿∗ ), the

former fund needs to have a cost advantage as follows: > −× ×

𝛥 𝑞 , where 𝛥𝑞 =

𝑞 − 𝑞 and 𝛥 𝑞 = − . A positive value of the ratio, , indicates a cost advantage of

fund 𝑗 + 1.

2) When the sampling probability is the same for all funds (i.e., 𝜌 = 1/𝑁) and when the distribution

density (i.e., 𝑔(𝑐)) is a constant, the cost advantage of fund 𝑗 + 1 can be simplified as >

> 0.

Proof: 1) We first note that 𝛿∗ < 𝛿∗ means 𝜙 − 2𝑏 𝑞 𝑍 < 𝜙 − 2𝑏 (𝑞 + Δ𝑞 )𝑍 + Δ 𝑞 .

Rearranging terms gives the expression stated in the first property.

2) When 𝜌 = 1/𝑁, equation (IA9) becomes 𝑞 = + ∑( )

( )( )−

( ), from which we can

get Δ𝑞 = 𝑞 − 𝑞 =( )( )

−×

( )+

( )= 𝐺 𝑐 − 𝐺 𝑐 > 0. Note

that 𝐺 𝑐 − 𝐺 𝑐 > 0 because 𝐺 𝑐 indicates that the distribution function of investors has lower than

𝑐 search costs.

Next, from equation (IA11), − =( )( )

+( )

𝑔 𝑐 −( )

𝑔 𝑐 =

( )𝑔 𝑐 −

( )

( )( )=

( )

( )( )𝑔 𝑐 − 𝑔 𝑐 , which gives Δ 𝑞 =

( )

( )( )𝑔 𝑐 − 𝑔 𝑐 . When 𝑔(𝑐) is a constant, Δ 𝑞 = 0. Hence, > > 0.

Proposition IA2 indicates that, since funds’ operation exhibits diseconomies of scale, funds with a

sufficiently large advantage in operational costs have an incentive to deliver higher performance in order

to attract more capital for their operation. The second part of the proposition illustrates an intuitive example,

in which we further assume that all funds have the same sampling probability and that the distribution

density is the same for all investors (e.g., a uniform distribution). Fund 𝑗 + 1 is willing to generate higher


performance (relative to fund 𝑗) when it has a relatively lower operational cost, even after accounting for

the additional capital attracted by its superior performance.9

Jointly, the two propositions and the search rule of equation (IA8) define the equilibrium of the fund

industry with search frictions, in which investors optimally search for funds and funds optimally deliver

performance to compensate for investors’ search costs. In particular, Propositions IA1 and IA2 indicate that

fund performance is jointly determined by the cost advantage of funds and the flow-performance sensitivity

of investors. Although there could be different ways of modeling the utilities of managers and the search

friction of investors, the model above provides a reasonable framework for understanding how capital and

search frictions from the investor side influence the incentives of fund managers to deliver performance.

B. Comparing to the Empirical Observations of Hedge Funds

We now apply the investor-search framework to hedge fund startups. We pay special attention to the

conditions under which the search framework can generate several fundamental properties of the hedge

fund industry. We show that a manager-search mechanism could arise for new funds in order to address the

severe frictions faced by these funds.

B.1. Performance Heterogeneity of Hedge Funds

We first state empirical observations. We then discuss related theoretical properties in a series of corollaries.

OBSERVATION IA1: Our empirical analysis features four types of funds in total: 1) existing funds, 2)

affiliated clone inceptions, 3) affiliated nonclone inceptions, and 4) stand-alone inceptions. Their

performance differs in our empirical observations.

CORROLARY IA1: Proposition IA1 is consistent with the existence of heterogeneous fund performance: 𝛿 ≤ 𝛿 _ < 𝛿 _ < 𝛿 _ .

Corollary IA1 provides an interpretation of hedge funds based on Proposition IA1, which allows for

performance heterogeneity of various types of hedge fund startups. Indeed, regardless of whether the search

costs come from the investor side or the manager side, the need to compensate heterogeneous search costs

9 The explicit functional form of the cost-benefit restriction arises from our assumptions on fund heterogeneity in operational costs. However, it is generally true that funds with cost advantages will choose to deliver higher performance and attract more capital in this type of search equilibrium. For instance, if we assume that managers have exactly the same operational costs but heterogerous learning costs, then the conclusion becomes funds with lower learning costs will deliver higher performance and attract more capital.


necessarily leads to performance heterogeneity. In this regard, investor-search and manager-search

frameworks generate equivalent results.

B.2. The Performance-size Relationship for Funds with the Same Sampling Probability

Further scrutiny of the performance-size relationship indicates that the two frameworks diverge

when we consider the key features of hedge funds startups. On the empirical side, we observe that inceptions

have both smaller size and better performance. In contrast, the investor-search model predicts a positive

size-performance relationship when funds have similar sampling probabilities.

OBSERVATION IA2: Inceptions have both superior performance to and smaller size than existing funds.

In particular, as summarized in Table IA1, the fund size for existing funds ($208 million) is much larger

than that of inceptions ($44.6 million) in our sample.

CORROLARY IA2: Higher performance attracts more capital when 𝜌 = 1/𝑁 for all funds.

Proof: Without the loss of generality, we examine the performance-size relationship for funds 𝑗 and 𝑗 + 1.

As shown in the second property of Proposition 2, when 𝜌 = 1/𝑁, 𝑞 − 𝑞 = 𝐺 𝑐 − 𝐺 𝑐 >

0.

The above corollary reveals an important property of the investor-search based equilibrium. For two

funds with a similar sampling probability, fund performance is positively correlated with fund size. The

intuition is as follows. When the sampling probability is the same, the better fund attracts not only the

capital from high-cost investors that the bad fund has access to, but also additional capital from low-cost

investors who actively search.

This property resembles the positive size-managerial skill relationship of Berk and Green (2004). This

consistency is not surprising. As we have discussed in Proposition IA1, a positive flow-performance

sensitivity remains in our search framework. This positive flow-performance sensitivity allows more capital

to be allocated to better-performing funds in equilibrium.

B.3. The Performance-size Relationship for Funds with Different Sampling Probabilities

The inconsistency between Corollary IA2 and Observation IA2 suggests that there are additional

sources of friction in the hedge fund industry that explain the performance of new funds. Within the


parameter set of our search framework, the friction is closely related to the sampling probability. Intuitively,

investors more easily find an existing fund than an inception (i.e., 𝜌 > 𝜌 ), which helps

explain the smaller size of the latter.

In this and the next subsection, we demonstrate two results related to the above intuition. First, we

confirm that a restriction on the sampling probability gives rise to a negative relationship between

performance and size. Second, when the sample probability of the inception is very small, however, as

implied by the data, it will incentivize fund managers to search for investors.

To more formally examine the implication of the heterogeneous sampling probability, we can examine

an economy in which funds 1,2, … , 𝑗 are existing funds and funds 𝑗 + 1, … , 𝑁 are inceptions. Assume that

all existing funds have the same sampling probability, which is higher than that of inception as (i.e., 𝜌 >

𝜌 ). In this case, the most visible differences in performance and size occur between fund 𝑗 + 1 and 𝑗.

Hence, we want to quantify the conditions under which fund sizes 𝑞 and 𝑞 match the empirically

observed fund size.

CORROLARY 3: Consider an economy in which funds 1,2, … , 𝑗 are existing funds and funds 𝑗 + 1, … , 𝑁 are inceptions and 𝜌 = ⋯ = 𝜌 > 𝜌 = ⋯ = 𝜌 . The following conditions hold in equilibrium:

1) 𝑞 = 𝑞 +⋯

𝐺 𝑐 − 𝐺 𝑐 , and

2) > .

Proof: From equation (IA9), we have − =⋯ ⋯

−⋯

+

⋯=

⋯. Rearranging terms proves the first condition. Next, notice that 𝐺 𝑐 −

𝐺 𝑐 > 0. Plugging this condition into the first property, we have > .

From Corollary IA3, inceptions do indeed have a smaller size than existing funds when the sampling

probability of the latter fund is sufficiently large. To generate the empirical observation that the size of a

typical existing fund is about 4.7 times as large as that of typical inception ($

$ . ≈ 4.7), we need

the difference in sampling probability to be even larger. For an economy with many existing funds, it will

be very difficult for investors to find a new fund through their sequential search process.

B.4. The Dual Impact of High Investors Search Frictions


If investors’ search capacity is so restrictive on new funds, it should exert a dual impact on these

funds. On one hand, it should incentivize managers of these funds to spend resources to find investors (i.e.,

managers start to search for investors). On the other hand, the influence of investors in shaping fund

performance declines because a smaller sampling probability reduces the flow-performance sensitivity. In

the extreme case, if investors cannot find a new fund at all, then the process of investor-search becomes

irrelevant to the fund—the new fund needs to find investors on its own. This dual impact can be more

formally demonstrated by the following corollary, which follows the setup of Corollary IA3.

COROLLARY IA4: 1) Denote 𝑈 = 𝑓 𝑞 𝑍 as the utility gain of the new manager. The marginal influence of the

sampling probability on the utility gain exceeds that of performance when the former probability is

below a certain threshold. Mathematically, > when 𝜌 < 𝛹 / , where

𝛹 𝜌 , … , 𝜌 , 𝐺(𝑐 ), … 𝐺 𝑐 ≡ 𝑞 /𝜌 indicates the part of the market share not directly

related to 𝜌 in equation (IA9).

2) A reduction in the sampling probability of the new fund 𝑗 + 1 reduces its flow-performance sensitivity,

that is., > 0. Moreover, when the sampling probability is very small (i.e., below a certain

threshold), this effect dominates fund performance—a reduction in the sampling probability also

reduces fund performance (i.e., ∗

> 0).

Proof: 1) Rewrite 𝑈 as 𝑈 = 𝜙 − 𝑏 𝑞 𝑍 𝑞 𝑍. We have = and =

. It follows that = 𝑍 𝜙 − 𝑏 𝑞 𝑍 , = 𝛹 , and = 𝜌 . Hence,

/ = 𝜌 /𝛹 , from which we derive the first property.

2) From equation (IA11), we have > 0. Hence, a reduction in 𝜌 leads to a decrease in

the flow-performance sensitivity. Next, since 𝛿∗ = 𝑓 × , we have ∗

= × +

𝑓 × = − 𝛹 + × , where the second equation

substitutes = − and = 𝛹 . Next, define 𝐹 ≡ / . We can then

see that ∗

> 0 iff 𝐹 < . Moreover, since is a quadratic function of 𝜌 , 𝐹 is an


increasing function of 𝜌 . In this case, let �̅� be the value of 𝜌 that solves 𝐹 = . Then

for any value of 𝜌 < �̅� , we have 𝐹 < and ∗

> 0.

To illustrate an intuitive example, we can consider an economy in which there is only one new fund

(i.e., 𝑁 = 𝑗 + 1). In this case, = 𝜌( )

+( )

( )( )+ ∑

( )

( ⋯ )( ⋯ ).

From this, we see that 𝐹 = 𝜌 /2, �̅� = . We have ∗

> 0 iff 𝜌 < �̅� . This

proves the second property.

Corollary IA4 proves that very high search frictions on the investor side (i.e., a very small sampling

probability in our model) can induce managers to search for investors and reduce the influence of investors

in shaping fund performance.

The first effect, as described in the first property of the corollary, describes what could be the missing

element in explaining the incentives of new funds in the investor-search framework. When the sampling

probability is lower than a certain threshold, its marginal influence on the utility gain of managers surpasses

that of performance. In this case, managers could be more interested in improving the sampling probability

than generating performance. For instance, managers may want to spend resources to search for investors,

which is equivalent to improving the sampling probability. Indeed, if the cost of the managerial search is

comparable to the cost of learning, then the manager should optimally choose to search.

The second effect is achieved through the reduced flow-performance sensitivity: as search frictions

from the investor side increase, capital flows are less sensitive to performance. In this case, managers have

a weaker incentive to deliver performance, following the intuition of Proposition 1. A highly restrictive

search process by investors over new funds, in this regard, is unlikely to explain the level of the superior

performance delivered by these new funds that we observe in the data.

Under the dual impact, therefore, the modeling of manager-initiated search becomes necessary to

complete the economic picture of how capital flows to new funds and subsequently influences their

performance when investors face very high search frictions. In our main text, we achieve this goal by

providing a manager-search framework, which complements the investor-search framework in describing

the incentives and performance of various types of inceptions.


C. A Comparison between Investor-Search and Manager-Search Frameworks

Corollary IA4 confirms the relevance of manager-search for hedge fund startups even when we start

from an investor-search framework. It also suggests that the invest-search framework documented here and

the manager-search framework described in our main text may capture different search incentives of the

hedge fund industry. To better understand the two frameworks, we provide a further comparison of their

major assumptions and implications.

C.1. Assumptions and the Degree of Freedom in Search-related Parameters

Aside from family structure considerations (which we do not study in our investor-search model), both

search frameworks involve three sets of main assumptions to determine the equilibrium distributions of

capital (fund size) and performance:

1) The BG condition and the tradeoff faced by fund managers;

2) Investor heterogeneity in supplying capital to funds;

3) Search rules that determine the allocation of capital to funds.

The first set of assumptions on managerial incentives are similar across the investor-search (hereafter

in this section, IS) and the manager-search (hereafter, MS) frameworks. In both cases, search costs are

compensated by fund performance, and fund managers face a tradeoff between the marginal benefit of

performance in increasing fund size and the marginal cost of learning. This tradeoff also resembles the

decision-making process of managers in BG.

The two frameworks diverge regarding the second and third sets of assumptions. On the investor side,

IS assumes that investors have heterogeneous search costs, whereas (two types of) MS investors differ in

their intrinsic valuation of the existing fund. Perhaps most importantly, search rules differ in the two

economies with different types of search frictions. In IS, the optimal rule is for investors to keep searching

until the expected marginal benefit of searching equals the marginal cost. In MS, the optimal rule is for

managers to search and bargain until the optimal fund size is achieved. We will see that the difference in

searching rules gives rise to important distinctions between the two economies.

Despite the differences in the second and third sets of assumptions, it is interesting to observe that the

two frameworks involve similar numbers of independent search parameters. When 𝑁 = 4, we have the

following set of parameters from the search side of IS: four sampling probability parameters, 𝜌 , four of

search cost parameters, 𝑐 , and a minimum of four parameters for 𝐺(𝑐 ). Since 𝑐 = 0 and all probabilities


and distribution densities sum to one, we have nine degrees of freedom in one equilibrium. When we further

treat clone inception as the same as the existing fund, then we have six independent search-related

parameters.

For MS, Table IA1 reports that we calibrate the manager-search model using five independent search-

related parameters, including search efficiency 𝑎, matching intensity 𝜌, bargaining efficiency 𝜉, and two

transition probabilities (𝜆 and 𝜆 ). Table IA1 also presents six steady-state parameters. But these

parameters are determined by the steady-state restrictions of the existing fund and do not provide additional

degrees of freedom.

C.2. Capital-Performance Sensitivity

The difference between the two search mechanisms is, therefore, not about the degree of freedom that

they offer. Rather, they diverge in two key implications of search rules related to the relationship between

capital and managerial skill and the split of economic rents between investors and funds.

The relationship between capital and managerial skills is widely observed in the mutual fund industry

as having a positive flow-performance sensitivity, which can be justified by the rational expectation

equilibrium of BG. The first-order relationship between capital and managerial skill is similar in IS: other

things being equal, more capital flows from investors to better-performing funds. The similarity between

IS and BG is not surprising. Since low-cost investors in IS can be regarded as BG investors, IS should

intuitively converge to BG as the fraction of investors with lower search costs increases.

The first-order effect in MS, however, is the substitution between the extensive margin and the intensive

margin from the perspective of fund managers. Compared to IS, this substitution effect implies the existence

of a reverse capital-performance relationship and reverse causality. That is, instead of capital being

determined by fund performance (IS), the performance incentive of the manager is determined by the

amount of capital she can find through her search (MS). Easier access to capital reduces the managerial

incentive to use performance as a bargaining tool to attract capital.

Conceptually, both capital-performance relationships may exist, as managers and investors should

respond to each other’s reactions. However, the reverse capital-performance relationship seems to play a

particularly important role in explaining the behavior of new funds because the sequence of actions in hedge

fund inceptions—i.e., to first raise capital and then deliver performance in future days, if not years—allows

the reverse causality to dominate the capital-performance relationship.


The reverse-causality relationship is not unique to hedge fund startups. In the private equity (PE)

industry, for instance, PE funds raising capital during the boom period of a business cycle tend to deliver

poorer performance (e.g., Harris, Jenkinson, and Kaplan, 2014). We believe that the MS model provides

the most suitable framework to explain managerial incentives in these cases.

C.3. The Split of Economic Rents

Another crucial difference between IS and MS lies in the distribution of the economic rents created by

fund managers.

In IS, just in BG, managers keep the economic rents they create. In BG, the main economic driver of

this result is that capital is supplied to funds competitively. The intuition of IS is similar: when investors

search—and therefore compete—for managerial skills, they get a competitive return of zero after search

costs. The competitiveness is embedded in the optimal search rule of investors. Since investors keep

searching until their expected marginal benefit equals their marginal cost, they get zero search-adjusted

economic rents.

For hedge fund startups, however, managers search (and thus compete) for accredited investors. An

intuitive result of this process is that investors retain a portion of the economic rents. This result also arises

from the optional search rule of managers: since a manager needs to search and bargain until the optimal

fund size is achieved, she is willing to share some economic rents with investors when this goal is otherwise

difficult to achieve. In this regard, we can also interpret the intensive margin of fund-raising as the optimal

strategy for the manager to share the economic rents with investors. Of course, this effect is stronger

(weaker) when the extensive margin of fund-raising is smaller (larger).

Again, we believe that both sharing schemes should exist in the economy. In practice, a typical active

mutual fund manager delivers below-market returns and retains most of the economic rents she creates. In

contrast, existing hedge funds deliver positive after-fee performance (about 2.72% per year as summarized

in Table IA1), on top of which new hedge funds deliver an additional alpha of approximately 2.84% per

year. In this regard, hedge fund managers share economic rents with investors and the sharing effect is more

prominent for new managers. MS, therefore, provides a suitable framework to study the managerial

incentives of sharing economic rents.

C.4. Toward a General Model of Two-way Search


The above discussion suggests that, as we often observe in practice, searches run in both directions. We

can view models with investor search frictions only and models with manager search frictions only as partial

equilibrium results of a general two-way search model.

Importantly, the explanatory power and the economic implications of the two kinds of search

frameworks are not the same. In particular, managerial search becomes important when investors are

restricted by very high search costs. Hedge fund inceptions have this property: practitioners widely view

raising money as among the most difficult business aspects of a new hedge fund (see footnote 2 in our main

text).

Moreover, hedge fund startups feature two important properties: a negative capital-performance

relationship and the splitting of economic rents between managers and investors. These properties are more

consistent with MS than IS.

Hence, in our main text, we focus on the manager-search framework to guide our understanding of the

economics of hedge fund startups. We view this model as a necessary complement to investor-search

models provided by the literature and a step toward a tractable two-way money-management search model.

D. Concluding Remarks

In this section, we present an extended Berk and Green (2004) model with investor-search frictions. In

linking this theoretical framework to our empirical observations, we notice that: 1) This type of market

friction generates performance heterogeneity across different types of funds. 2) Funds with better

performance can attract more capital when funds are associated with similar sampling probabilities from

investors’ perspective. This general property, achieved through positive flow-performance sensitivity,

resembles the positive size-managerial skill relationship found in Berk and Green (2004). 3) A

heterogeneous sampling probability can help explain the reverse size-performance relationship that hedge

fund startups exhibit (i.e., hedge fund startups have both superior performance and a much smaller size than

existing funds).

However, to match the empirically observed size difference between new and existing hedge funds, our

investor-search model indicates that investors would need to face extremely high search frictions when

finding new funds. In this case, fund performance would be influenced less by investor choice (due to

reduced flow-performance sensitivity) and more by manager actions (because of the tremendous benefits

these funds enjoy when actively reaching out to investors). These considerations give rise to manager-

initiated search, which better predicts several empirical observations noted in our main text.


We further show that investor-search and manager-search frameworks diverge in two key implications

of their search mechanisms: the relationship between capital and managerial skill, and the split of economic

rents between investors and funds. Although the manager-search framework better fits empirical

observations, we believe that a two-way money-management search model could provide additional

insights as it synchronizes the relationship between capital and managerial skills. We leave the task of

constructing a tractable two-way search model for future research.


REFERENCES

Agarwal, Vikas, Y. Eser Arisoy, and Narayan Y. Naik, 2017, Volatility of aggregate volatility and hedge fund returns, Journal of Financial Economics 125, 491-510.

Agarwal, Vikas, Vyacheslav Fos, and Wei Jiang, 2013, Inferring reporting-related biases in hedge fund databases from hedge fund equity holdings, Management Science, 51, 1271-1289.

Amihud, Yakov and Ruslan Goyenko, 2013, Mutual Fund’s R2 as Predictor of Performance, Review of Financial Studies 26, 667-694.

Bali, Turan G., Stephen J. Brown, and Mustafa O. Caglayan, 2014, Macroeconomic risk and hedge fund returns, Journal of Financial Economics, 114, 1-19.

Berk, Jonathan B. and Richard C. Green, 2004, Mutual fund flows and performance in rational markets, Journal of Political Economy, 112, 1269–1295.

Berk, Jonathan B. and Jules H. van Binsbergen, 2015, Measuring skill in the mutual fund industry, Journal of Financial Economics, 118, 1–20.

Berk, Jonathan B. and Jules H. van Binsbergen, 2017, Mutual funds in equilibrium, Annual Review of Financial Economics, 9, 147–167.

Buraschi, Andrea, Robert Kosowski, and Fabio Trojani, 2014, When there is no place to hide: Correlation risk and the cross-section of hedge fund returns, Review of Financial Studies, 27, 581-616.

Duffie, Darrell, Nicolae Gârleanu, and Lasse Heje Pedersen, 2007, Valuation in over-the-counter markets. Review of Financial Studies, 20, 1865-1900.

Fung, William and David Hsieh, 1997, Empirical characteristics of dynamic trading strategies: The case of hedge funds. Review of Financial Studies, 10, 275–302.

Fung, William and David Hsieh, 2000, Performance characteristics of hedge funds and commodity funds: Natural vs. spurious biases, Journal of Financial and Quantitative Analysis, 35, 291-307.

Fung, William and David A. Hsieh, 2004, Hedge fund benchmarks: A risk-based approach, Financial Analysts Journal, 60, 65–80.

Gârleanu, Nicolae and lasse Heje Pedersen, 2018, Efficiently inefficient markets for assets and asset management, Journal of Finance, 73, 1663–1712.

Getmansky, Mila, Andrew W. Lo, and Igor Makarov, 200, An econometric model of serial correlation and illiquidity in hedge fund returns, Journal of Financial Economics, 74, 529–609.

Glode, Vincent and Richard C. Green, 2011, Information spillovers and performance persistence for hedge funds, Journal of Financial Economics, 101, 1–17.

Harris, Robert S., Tim Jenkinson, and Steven N. Kaplan, 2014, Private equity performance: What do we know?, Journal of Finance, 69, 1851–1882.

Hortaçsu, Ali, Chad Syverson, 2004, Product differentiation, search costs and competition in the mutual fund industry: A case study of S&P 500 index funds, Quarterly Journal of Economics, 119, 303-356.

Jagannathan, Ravi, Alexey Malakhov, and Dmitry Novikov, 2010, Do hot hands exist among hedge fund managers? An empirical evaluation, Journal of Finance, 65, 217–255.

Pástor, Lubos and Robert F. Stambaugh, 2003, Liquidity Risk and Expected Stock Returns, Journal of Political Economy, 111, 642-685.

Rubinstein, Ariel and Asher Wolinsky, 1985, Equilibrium in a market with sequential bargaining, Econometrica, 53, 1133–50.


Sadka, Ronnie, 2010, Liquidity risk and the cross-section of hedge-fund returns, Journal of Financial Economics, 98, 54-71.

Sun, Zheng, Ashley Wang, and Lu Zheng, 2012, The road less traveled: strategy distinctiveness and hedge fund performance, Review of Financial Studies, 25, 96–143.


III. Data Description

Table IAII Fund Counts by Database

This table reports the counts of funds in our merged universe by database of origin. The three databases are Lipper TASS, BarclayHedge, and HFR. Fund name and return information are used to identify overlapping funds using a combination of computer and manual techniques. The first panel provides the number of qualifying funds in each database before our merge procedure. The second panel indicates how many funds in the merged sample were taken from each database. When a fund appears in more than one database, we use the prioritization TASS, HFR, BarclayHedge to determine from which database we draw fund information. We further provide a Venn diagram of sample funds according to which database(s) they report to. These counts are logically prior to the imposition of filters on the number of available returns in each fund.

Group Funds

Total funds in TASS Database 11,668 Total funds in HFR Database 17,425 Total funds in BarclayHedge Database 17,309

TASS funds used in our merged sample 11,668 HFR funds used in our merged sample 11,534 BarclayHedge funds used in our merged sample 8,200


Table AIII Strategy Mapping Table

This table presents an example of strategy mapping across different hedge fund databases. For each strategy or strategy-substrategy reported in BarclayHedge or HFR, we identify the corresponding strategy in TASS with the greatest overlap of known matching funds. Funds from HFR or BarclayHedge are mapped into a TASS strategy according to the resulting mapping. TASS Strategy BarclayHedge Strategy HFR Strategy Convertible Arbitrage Convertible Arbitrage

Convertible Arbitrage - Credit Convertible Arbitrage - Volatility Fixed Income - Convertible Bonds

Relative Value/Fixed Income - Convertible Arbitrage

Dedicated Short Bias Equity Dedicated Short Equity Short-Bias Equity Short-Bias - Growth Oriented Equity Short-Bias - Opportunistic

Equity Hedge/Short Bias

Emerging Markets Emerging Markets - Asia Emerging Markets - Eastern Europe/CIS Emerging Markets - Global Emerging Markets - Latin America Emerging Markets - MENA Emerging Markets - Other

Equity Market Neutral Equity Market Neutral Equity Market Neutral - Quantitative Equity Market Neutral - Value Oriented Statistical Arbitrage

Equity Hedge/Equity Market Neutral

Event-Driven Distressed Securities Event-Driven Merger Arbitrage

Event-Driven/Distressed/Restructuring Event-Driven/Merger Arbitrage Event-Driven/Multi-Strategy Event-Driven/Private Issue/Regulation D Event-Driven/Special Situations


Fixed Income Arbitrage Fixed Income - Arbitrage Fixed Income - Diversified Fixed Income - High Yield Fixed Income - Long/Short Credit Fixed Income - Mortgage Backed Fixed Income - Unconstrained

Relative Value/Fixed Income - Asset-Backed Relative Value/Fixed Income - Corporate Relative Value/Fixed Income - Sovereign

Global Macro Discretionary Fundamental - Currency Fundamental - Diversified Fundamental - Financial/Metals Macro Macro - Discretionary Tail Risk Technical - Currency

Macro/Currency - Discretionary Macro/Currency - Systematic Macro/Discretionary Thematic Macro/Multi-Strategy


Long/Short Equity Hedge Activist Balanced (Stocks & Bonds) Closed-end funds Equity 130-30 Equity 130-30 - Growth Oriented Equity Long Only Equity Long Only - Growth Oriented Equity Long Only - Opportunistic Equity Long Only - Quantitative Equity Long Only - Trading Oriented Equity Long Only - Value Oriented Equity Long-Bias Equity Long-Bias - Growth Oriented Equity Long-Bias - Opportunistic Equity Long-Bias - Quantitative Equity Long-Bias - Trading Oriented Equity Long-Bias - Value Oriented Equity Long/Short Equity Long/Short - Growth Oriented Equity Long/Short - Opportunistic Equity Long/Short - Quantitative Equity Long/Short - Trading Oriented Equity Long/Short - Value Oriented Fundamental - Energy Sector - Energy Sector - Environment Sector - Farming Sector - Financial Sector - Health Care/Biotech Sector - Metals/Mining Sector - Miscellaneous Sector - Natural Resources Sector - Real Estate Sector - Technology

Equity Hedge/Fundamental Growth Equity Hedge/Fundamental Value Equity Hedge/Multi-Strategy Equity Hedge/Quantitative Directional Equity Hedge/Sector - Energy/Basic Materials Equity Hedge/Sector - Healthcare Equity Hedge/Sector - Technology Event-Driven/Activist Macro/Commodity - Metals


Managed Futures Arbitrage Single Advisor Fund Stock Index Systematic Technical - Diversified Technical - Energy Technical - Financial/Metals Technical - Interest Rates

Macro/Commodity - Agriculture Macro/Commodity - Energy Macro/Commodity - Multi Macro/Systematic Diversified

Multi-Strategy Fundamental - Agricultural Macro - Quantitative Multi-Strategy Multi-Strategy -Strategy Arb - Quantitative PIPEs Volatility Trading

Relative Value/Multi-Strategy


Table IAIV Summary Statistics of Hedge Funds and Hedge Fund Families

This table reports the number of funds in our sample by investment strategy category and year. Funds are included if they report a return in December of the given year and have at least 12 monthly return observations. Strategy classifications are convertible arbitrage (CA), dedicated short bias (DS), event-driven (ED), emerging markets (EM), equity market neutral (EMN), fixed income arbitrage (FI), global macro (GM), long/short equity (LS), managed futures (MF), and multi-strategy (MS). We exclude funds of funds and funds without strategy information. We also report the total number of hedge funds and hedge fund families, as well as the average number of funds per family and the fraction of families that have multiple funds at the end of each year. Our sample is formed by merging fund/family/return information from TASS, HFR, and BarclayHedge and spans the period 1994 to 2016.

Strategy Category Counts

Year CA DS ED EM EMN FI GM LS MF MS Number of

Funds Number of

Families Funds per

Family Fraction

Multi-Fund 1994 77 25 180 81 72 83 297 781 886 110 2,592 1,555 1.67 0.30 1995 91 23 234 115 89 91 338 960 926 129 2,996 1,722 1.74 0.33 1996 106 29 285 143 98 115 320 1,195 898 147 3,336 1,890 1.77 0.34 1997 114 33 338 188 136 131 328 1,391 887 178 3,724 2,052 1.81 0.36 1998 135 35 373 194 179 146 339 1,599 835 197 4,032 2,215 1.82 0.36 1999 153 44 423 218 217 180 330 1,892 768 240 4,465 2,374 1.88 0.37 2000 190 48 482 226 238 204 340 2,154 732 271 4,885 2,497 1.96 0.39 2001 239 44 550 219 321 259 375 2,476 726 318 5,527 2,711 2.04 0.40 2002 289 45 624 242 409 325 456 2,706 750 413 6,259 2,883 2.17 0.41 2003 320 48 694 288 459 442 609 3,131 811 543 7,345 3,210 2.29 0.43 2004 312 45 768 375 513 538 734 3,616 921 713 8,535 3,512 2.43 0.44 2005 294 46 825 499 578 605 840 4,128 987 853 9,655 3,835 2.52 0.44 2006 280 48 872 646 595 636 936 4,558 1,038 1,003 10,612 4,042 2.63 0.44 2007 239 52 881 782 601 630 991 4,742 1,085 1,193 11,196 4,167 2.69 0.44 2008 178 44 743 808 544 605 1,015 4,451 1,126 1,251 10,765 4,070 2.64 0.43 2009 156 42 743 847 532 661 1,140 4,494 1,200 1,284 11,099 4,109 2.70 0.42 2010 180 39 778 881 494 764 1,194 4,602 1,282 1,397 11,611 4,124 2.82 0.43 2011 190 34 768 882 493 828 1,208 4,683 1,325 1,389 11,800 4,015 2.94 0.44 2012 200 20 719 872 464 866 1,186 4,631 1,322 1,274 11,554 3,877 2.98 0.43 2013 206 14 726 733 439 829 1,016 4,651 1,252 1,100 10,966 3,743 2.93 0.42 2014 196 19 701 671 462 848 924 4,485 1,158 946 10,410 3,526 2.95 0.42 2015 172 11 629 625 439 796 792 4,141 1,014 799 9,418 3,205 2.94 0.41 2016 159 12 547 568 379 695 653 3,583 779 690 8,065 2,812 2.87 0.40


IV. Details on Market-Timing & Security-Selection Bootstrapping Method

Our bootstrap approach involves randomly resampling the regression residuals of equation (7) in the

main article to generate pseudo-portfolios that have the same factor loadings as the actual portfolio p, but

have no timing ability (by construction). We then assess whether the estimated coefficient of interest for

the inception portfolio has a p-value less than 5% based on the bootstrapped distribution that assumes no

timing ability. Our bootstrap procedure is illustrated for the case 𝛾 = 0 as follows:

i. Estimate equation (7) in the main article for a given inception portfolio and save the estimated

coefficients 𝛼 , �̂� , 𝛽 , , … , 𝛽 , and the time series of residuals 𝜖̂ , ; 𝑡 = 1, 2, … , 𝑇 .

ii. Generate a randomly resampled residual time series 𝜖̂ , ; 𝑡 = 1, 2, … , 𝑇 by sampling the

residuals with replacement. Here, b is the index of bootstrap iterations, {𝑏 = 1, 2, … , 𝐵}. Next, we

obtain the monthly excess returns for a pseudo inception portfolio that has no timing ability by

setting 𝛾 = 0:

𝑟 , = 𝛼 + 𝛽 , 𝑀𝑘𝑡 + 𝛽 , 𝑆𝑀𝐵 + 𝛽 , 𝑌𝐿𝐷𝐶𝐻𝐺 + 𝛽 , 𝐵𝐴𝐴𝑀𝑇𝑆𝑌

+ 𝛽 , 𝑃𝑇𝐹𝑆𝐵𝐷 + 𝛽 , 𝑃𝑇𝐹𝑆𝐹𝑋 + 𝛽 , 𝑃𝑇𝐹𝑆𝐶𝑂𝑀 . (IA14)

iii. Estimate the timing model of equation (7) from the main article using the pseudo portfolio returns

from equation (IA14) and save the estimate of the timing coefficient. By construction, the pseudo-

inception portfolio has 𝛾 = 0, so any nonzero coefficient is attributable to the sampling errors.

iv. Repeat the above steps for 10,000 iterations to generate the empirical distribution of 𝛾 (i.e.,

B=10,000). Then compute the empirical p-value of 𝛾 as the proportion of the values of the

bootstrapped 𝛾 for the pseudo-inception portfolios that exceed the estimated 𝑟 for the actual

inception portfolio.


V. Robustness Checks

In this part of the Internet Appendix, we provide detailed discussions of and results for the

robustness tests mentioned in the main text.

A. Return-smoothing

In the main text, we examine a leading alternative explanation for the performance difference

between cold and hot inceptions: exposure to illiquidity or return-smoothing. To address this

concern, we follow Getmansky, Lo, and Makarov (2004 hereafter GLM) to assess the serial

correlation associated with illiquidity and return-smoothing. For each fund, we follow GLM to

perform a moving average (MA) regression on excess returns,

𝑟 , = 𝜃 , 𝜈 , + 𝜃 , 𝜈 , + 𝜃 , 𝜈 , ,

where 𝑟 , is the observed excess return for fund 𝑖 in month 𝑡, and 𝜈 , is its true economic return.

We follow GLM to focus only on the estimators that yield invertible MA processes and require 60

observations for estimation1. The maximum likelihood estimation imposes the following

normalization: 𝜃 , + 𝜃 , + 𝜃 , = 1. The degree of serial correlation caused by illiquidity and

return-smoothing can be captured by two parameters: the contemporaneous MA parameter, 𝜃 , ,

and the smoothing index 𝜉 = 𝜃 , + 𝜃 , + 𝜃 , . If there is no return-smoothing or exposure to

illiquidity, fund residuals should follow a random walk and both 𝜃 , and 𝜉 should be equal to one.

In practice, the two parameters deviate from one—the more deviations, the more potential concern

that illiquidity and return-smoothing artificially inflate fund performance.

Table IAIV reports our estimation results. Panel A tabulates the average value of the MA

coefficients and the smoothing index for cold and hot inceptions. We first notice that both 𝜃 , and

𝜉 of cold and hot funds are smaller than one. Hence both types of inceptions exhibit return-

smoothing. Second, hot inceptions have significantly smaller θ and 𝜉 than cold inceptions do,

suggesting that the returns of hot inceptions exhibit more illiquidity and smoothing. The average

𝜃 coefficient is 0.891 and 0.936 for hot and cold inceptions, respectively, and the coefficient

1 We follow GLM to require 60 return observations but we also validate that inference is the same using a threshold of 48 or 36 observations.


difference between them is significant (with a t-statistic of 3.37). Similarly, the average smoothing

index, 𝜉, of cold funds is also significantly larger than that of hot funds, suggesting that the hot

inceptions exhibit more return-smoothing than cold inceptions (the difference is 0.082 with a t-

statistic of 2.77).

In Panel B, we confirm this difference by performing a cross-sectional regression of

smoothing measures on inception types. In this test, we pool all inceptions together (including

cold, hot, and others). The regression equation is

θ , or ξ = 𝑎 + 𝑏 × 𝐷 , + 𝑏 × 𝐷 , + 𝜂 ,

where 𝐷 , and 𝐷 , are dummy variables indicating whether inception i is classified as hot or

cold. Following GLM, we regress 𝜃 , and 𝜉 on dummy variables and control for the strategy-

level fixed effects. We find that the coefficient on 𝐷 is significantly positive, suggesting that

cold inceptions exhibit less return-smoothing than other inceptions. This observation is robust

across all specifications. To see if cold inceptions also exhibit less return-smoothing than hot

funds, we conduct an F-test on the null hypothesis of 𝑏 = 𝑏 . The last line of Panel B reports the

p-value of the test, which rejects the null that the two types of inceptions have the same return-

smoothing profile. Hence, the superior performance of cold over hot inceptions is unlikely to be

caused by return-smoothing.2

B. Alternative Risk Factors

In the main text, we mention that, in addition to the seven-factor model of Fung and Hsieh

(1997), funds can also be exposed to additional risk factors related to liquidity (Pastor and

Stambaugh (2003) and Sadka (2010)), correlation risk (Buraschi, Kosowski, and Trojani (2014)),

economic uncertainty (Bali, Brown, and Caglayan (2014)), and volatility-of-volatility (Agarwal,

Arisoy, and Naik (2017)). Between the two liquidity factors, Pastor and Stambaugh (2003) capture

market-wide liquidity based on daily price reversal, while Sadka (2010) focuses on the component

of liquidity associated with informed trading. In addition, the correlation risk factor of Buraschi,

Kosowski, and Trojani (2014) captures unexpected changes in asset correlation that may adversely

2 GLM also extend their test to allow the contemporaneous and two-period lagged values of a common factor to affect fund returns. We conduct an additional test to follow this idea as well. We also use the Fung and Hsieh (2004) seven-factor model to estimate alphas. When including the seven risk factors and lagged market excess returns in estimating alphas, we lose funds with shorter return history. Overall, we find that the conclusions are very similar.


influence portfolio diversification, the economic uncertainty factor of Bali, Brown, and Caglayan

(2014) synchronizes the time-varying conditional volatility of macroeconomic variables

associated with business cycle conditions, and the volatility-of-volatility factor of Agarwal,

Arisoy, and Naik (2017) measures uncertainty about the expected return on the market portfolio.

Since the volatility-of-volatility factor is available over a much shorter period (the data start in

2006), we focus our discussion on the first four recently proposed risk factors in this test.3

In Table IAV, we start from the seven-factor model explaining the alpha of the spread portfolio

between cold stand-alone and hot clone inceptions (i.e., column (8) in Table V) and add additional

factors one at a time. Except for the liquidity factor of Sadka (2010), the additional factors do not

have significant power in explaining the inception portfolio return spread. Sadka’s liquidity factor

is likely relevant because it is constructed from variables related to informed trading. Importantly,

however, the inclusion of the Sadka liquidity factor does not change our conclusions from Tables

V and VI. The coefficients on the risk-adjusted spread (alphas) are close to that reported in column

(8) of Table V. We conclude that our results are not explained by alternative risk factors.

C. Controlling for Fund Characteristics and Policy Choices

In the main text, we note that we conduct cross-sectional analysis to control for characteristics

of fund returns, including two measures of risk (market beta, return volatility), two measures of

how distinctive a fund’s strategy is with respect to other hedge funds in the same strategy category

or with respect to common risk factors in the market (the Strategy Distinctiveness Index, SDI, of

Sun, Wang, and Zheng (2012) and R2 of Amihud and Goyenko (2013)), as well as operational

policy choices related to incentive fees, management fees, and the redemption notice period and

redemption frequency (we express redemption frequency in days—a larger value indicates a more

restrictive redemption policy).

The more detailed specification in Table IAVI is as follows: for each inception i, we perform

the Fung and Hsieh (2004) seven-factor regression to obtain fund-specific alpha, 𝑎 during each

fund’s 60-month inception period. We then regress fund alpha on the dummy variable 𝐷 , ,

which takes the value of one if i is a cold stand-alone inception and zero if it is a hot clone

3 Over the short period when the volatility-of-volatility factor is available, we find that the return spread between cold stand-alone and hot clone inceptions does not have significant exposure to this risk factor.


inception, and on a vector of inception characteristics, 𝐹 , as follows:

𝛼 = 𝑎 + 𝑏 × 𝐷 , + 𝜃𝐹 + 𝜈 . (IA15)

In specification (1), we present a benchmark case to validate our previous findings without

controlling for fund characteristics and policies. We see that the cold stand-alone dummy is

associated with a significant and positive coefficient in this cross-sectional regression (0.324% per

month with a t-statistic of 7.05), which implies that cold stand-alone inceptions significantly

outperform hot clone inceptions.

In specification (2), we add alternative return characteristics measured at the fund level. For

instance, Sun, Wang, and Zheng’s (2012) Strategy Distinctiveness Index (SDI), a measure of how

unique a hedge fund’s returns are compared with other funds, is positively related to fund alpha.

In contrast, market beta is negatively related to risk-adjusted performance. This negative relation

is consistent with Table IV, in which the more skillful cold inceptions exhibit lower market beta

(0.202) than hot inceptions (0.320). With controls for these variables, our main result remains

unchanged (cold stand-alone inceptions are associated with higher alphas). Hence, the difference

between cold and hot inceptions is distinct from what is captured by fund characteristics.

In specification (3), we add endogenous policy choices. We find that fund policies related to

incentive fees, management fees, redemption notice, and redemption period affect alpha. For

instance, the coefficient on incentive fees is negative and significant, suggesting that charging high

incentive fees reduces the risk-adjusted return that investors receive. Also, a more restrictive

redemption notice period is associated with better performance. The inclusion of fund policy

controls does not affect the significance of the cold inception dummy.

Specification (4) adds the impact of fund flows on our results. We focus on an important

question related to our analysis: do inflows received by young hedge funds dilute their

performance incentives in the post-inception period? To address this question, we include fund

flows received by cold stand-alone inceptions and hot clone inceptions during their 60-month

inception period. For this test, fund flow calculated in month t from equation (4) is normalized by

its lagged AUM times (1 + 𝑟 , ), the counterfactual AUM of the fund without any flow. We require

a minimum of 24 months of valid information to compute average fund flows.4 The coefficient on

the cold inception dummy has a t-statistic of 4.34. This result suggests that it is the investor demand

4 We use a dummy variable to absorb the potential performance implication of funds that do not reach this minimum flow request.


that managers encounter during their capital-raising process (as opposed to fund flows that they

might attract after inception) that affects the performance incentives of new funds.

D. Alternative Holding Periods and Definitions of Cold and Hot inceptions

Next, we re-run Table VI using a shorter holding period of 48 months (Panels A1 and B1 of

Table AIV) and a longer period of 72 months (Panels A2 and B2). We find that the main features

of Table VI remain unchanged.

We also examine the impact of alternative definitions of cold and hot inceptions. A strategy is

classified as hot (cold) if its normalized inceptions are among the top (bottom) 30% of all strategies

over the 36 months prior to inception. Inceptions are normalized by dividing by the number of

funds in that strategy at the beginning of that period. We form inception portfolios based on both

family structure and this alternative strategy identification. The results are reported in Table AVIII.

As can be seen, our main conclusions from Table VI remain largely unchanged.

E. More Discussions on the Impact of Data Biases

In the main text, we discuss how we mitigate survivorship bias. In this part of the Internet

Appendix, we provide more details on the bias originated from voluntary reporting to databases.

Reporting to a database is voluntary and frequently done to attract investors. Both poor-

performing funds and star funds may choose not to report to any databases, creating a selection

bias. This bias is difficult to address because (1) we do not observe hedge funds that choose not to

report to any commercial database, and (2) we do not observe fund-level holdings in 13F filing

data, as the SEC requires institutional investors to report equity holdings only at the management

company (family) level. In contrast, mutual funds must disclose fund-level holdings. Moreover,

only hedge funds that hold more than $100M of U.S. securities and that use the U.S. postal service

are required to file 13F holdings. These important reporting differences make analysis of hedge

fund holdings much more difficult than that of mutual funds.

Based on anecdotal evidence, Fung and Hsieh (1997, 2000) conclude that the impact of

selection bias could be limited—the upward bias on returns resulting from missing information for

inferior funds is offset by the downward bias caused by star funds with superior performance


choosing not to disclose their information because they have no incentive to do so. Agarwal, Fos,

and Jiang (2013) use 13F filing data to examine the impact of selection bias at the hedge fund

company level and find no performance difference between self-reporting funds and nonreporting

funds, supporting the conclusion that the performance bias caused by selection is small.


Table AIV Tests of Return-smoothing

This table presents results of return-smoothing analysis for inceptions. Following Getmansky, Lo, and Makarov (2004), we investigate the serial correlation properties of hedge fund inceptions by performing the following moving-average regression to determine the degree of serial correlation for each inception:

𝑟 , = 𝜃 , 𝜈 , + 𝜃 , 𝜈 , + 𝜃 , 𝜈 , ,

where 𝑟 , is the observed excess return for fund 𝑖 in month 𝑡, and 𝜈 , is the true economic return of that

fund. Also, following Getmanskyn Lo, and Makarov (2004), we estimate via maximum likelihood, imposing the normalization 𝜃 + 𝜃 + 𝜃 = 1 in the above regression, and construct the smoothing

measure for fund 𝑖 as 𝜉 = 𝜃 , + 𝜃 , + 𝜃 , . Panel A tabulates the mean values of 𝜃 , (τ ∈ {0, 1, 2}) and 𝜉

for cold stand-alone inceptions and hot clone inceptions. In Panel B, we pool the sample of all inceptions (cold, hot, and other inceptions), and present the results of the following cross-sectional regression analysis:

θ , or ξ = a + b × D , + b × D , + η ,

where 𝐷 , and 𝐷 , are dummy variables that take the value of one for cold and hot inceptions,

respectively. We include fixed effects for the fund’s strategy. The last line of Panel B reports the p-value of the F-test for the null hypothesis of 𝑏 = 𝑏 . A minimum of 60 observations is required for inclusion. t-statistics are in parentheses. Statistical significance at the 1%, 5%, and 10% levels is denoted by ***, **, and *, respectively.

Panel A: Average Coefficients by Inception Type

𝜃 𝜃 𝜃 𝜉

Cold Stand-Alone Inceptions 0.936 0.054 0.01 0.97

Hot Clone Inceptions 0.891 0.089 0.02 0.888

Cold Stand-Alone minus Hot Clone Spread 0.044 -0.035 -0.01 0.082

(t-statistic) (3.365) (-3.816) (-1.135) (2.772) Panel B: Regression of Coefficients from Step 1 on Grouping

𝜃 𝜉 Cold Stand-Alone Inceptions 0.03*** 0.06***

(3.00) (3.01)

Hot Clone Inceptions -0.02** -0.03

(-2.10) (-1.62)

Fixed Effect: Strategy Yes Yes

Adj. R2 12.90% 9.00%

p-value of 𝑏 = 𝑏 0.000 0.000


Table IAV Regression Analysis of the Cold-Hot Inception Portfolio Return Spread on Additional Risk Factors This table presents results of the regression analysis for cold-hot inception return spread. We regress the return spread between the cold stand-alone inception portfolio and the hot clone inception portfolio on the Fung and Hsieh (2004) seven factors, augmented with additional risk factors. The regression equation is

𝑅𝑒𝑡𝑆𝑝𝑟𝑒𝑎𝑑 = 𝛼 + 𝛽 𝑀𝐾𝑇 + 𝛽 𝑆𝑀𝐵 + 𝛽 𝑌𝐿𝐷𝐶𝐻𝐺 + 𝛽 𝐵𝐴𝐴𝑀𝑇𝑆𝑌 + 𝛽 𝑃𝑇𝐹𝑆𝐵𝐷 + 𝛽 𝑃𝑇𝐹𝑆𝐹𝑋 + 𝛽 𝑃𝑇𝐹𝑆𝐶𝑂𝑀 + 𝛽 𝐹 + 𝜀 ,

where RetSpreatt is the return spread between the cold stand-alone inception portfolio and the hot clone inception portfolio in month t. The Fung-Hsieh (2004) factors are listed in Table IV. The factor F is one of the following: the Pastor Stambaugh (2003) liquidity factor, the Sadka (2003) liquidity factor, the Economic Uncertainty Index of Bali, Brown, and Caglayan (2014), the CRP correlation risk factor of Buraschi, Kosowski, and Trojani (2014), and the lagged market return. t-statistics are in parentheses. Statistical significance at the 1%, 5%, and 10% levels is denoted by ***, **, and *, respectively.


(1) (2) (3) (4) (5) Risk-adjusted Spread 0.531 *** 0.585 *** 0.532 *** 0.472 *** 0.546 ***

(4.65) (4.50) (4.64) (3.72) (4.66) Pastor Stambaugh Liquidity -0.017

(-0.90) Sadka Liquidity -0.727 *** (-3.17)

Economic Uncertainty Index 0.000 (0.02)

CRP Correlation Risk -0.011 (-1.11)

Lag(Mkt) -0.032 (-1.06)

Mkt -0.028 -0.041 -0.034 -0.042 -0.034 (-0.97) (-1.33) (-1.19) (-1.40) (-1.22)

SMB 0.094 *** 0.094 ** 0.095 *** 0.094 *** 0.095 *** (2.68) (2.45) (2.71) (2.85) (2.70)

YLDCHG 0.266 -0.026 0.189 0.332 0.093 (0.45) (-0.04) (0.32) (0.58) (0.16)

BAAMTSY 0.990 0.072 1.022 1.015 0.668 (1.33) (0.08) (1.35) (1.46) (0.81)

PTFSBD 0.025 *** 0.028 *** 0.024 *** 0.023 *** 0.021 *** (3.07) (3.15) (3.03) (2.85) (2.64) PTFSFX -0.016 ** -0.019 ** -0.016 ** 0.001 -0.015 **

(-2.42) (-2.58) (-2.39) (0.12) (-2.24) PTFSCOM 0.000 0.004 -0.001 -0.003 0.001

(-0.06) (0.43) (-0.06) (-0.42) (0.10)

Adj. R2 5.20% 9.60% 4.90% 5.60% 4.70%


Table IAVI Regression Analysis of Inception Alphas on Fund Characteristics and Policy Choices

This table reports results of cross-sectional regressions of inception alphas on fund policy choices and return characteristics, using cold stand-alone and hot clone inceptions. For each inception, we use the Fung-Hsieh (2004) seven-factor model to estimate its abnormal performance (alpha) during the 60-month period after the inception. These alphas are regressed on fund characteristics and policy choices as follows:

𝛼 = 𝑎 + 𝑏 × 𝐷 , + 𝜃𝐹 + 𝜈 ,

where 𝐷 , is a dummy variable equal to one if fund 𝑖 is a cold stand-alone inception and zero if it is a hot clone inception, and 𝐹 is a vector of fund return characteristics and fee/liquidity policies. For each fund i, its characteristics are as follows: 𝑆𝐷𝐼i is the strategy distinctiveness distance measure of Sun, Wang, and Zheng (2012), Mkt Beta is 𝛽 , from the first-stage regression, Volatility is the volatility of fund excess returns (r , ), and R Squaredi is the R2 from the Fund and Hsieh (2004) seven-factor regression. Incentive Fees and Management Fees are fees charged by the fund, Redemption Notice Period is the number of days’ notice required before making a redemption, and Redemption Frequency is the number of days between redemption windows. Fund flows are measured over the same holding period used to compute fund alpha. t-statistics are in parentheses. Statistical significance at the 1%, 5%, and 10% levels is denoted by ***, **, and *, respectively.

(1) (2) (3) (4)

Intercept 0.1113*** -0.1669 -0.0955 -0.0391

(3.46) (-1.64) (-0.73) (-0.29)

Cold Inception Dummy 0.3242*** 0.2399*** 0.2154 *** 0.2068***

(7.05) (5.45) (4.58) (4.34)

SDI 0.5196*** 0.4452 *** 0.4478***

(4.14) (3.36) (3.38)

Mkt Beta -0.7197*** -0.8198 *** -0.8129***

(-10.70) (-10.69) (-10.63)

Volatility 2.6316*** 2.8189 *** 2.8887***

(10.60) (9.66) (9.88)

R Squared from Fung Hsieh -0.1147 -0.2112 -0.2263

(-0.76) (-1.35) (-1.45)

Incentive Fee -1.4666 *** -1.5242***

(-4.04) (-4.19)

Management Fee 3.6282 3.6219

(0.82) (0.82)

Redemption Notice Period (days) 0.0069 *** 0.0066***

(7.47) (7.04)

Redemption Frequency (days) -0.0002 -0.0002

(-0.46) (-0.35)

Fund Flows 0.0746**

(2.47)

Adj. R2 2.60% 15.30% 17.10% 17.50%


Table AVII Strategy Demand and Family Structure: Evidence using alternative Holding Periods

This table presents a two-way summary of the risk-adjusted returns (alphas) of different types of inception portfolios using a 48-month (Panels A1 to B1) and a 72-month (Panels A2 to B2) holding period. Each portfolio’s alpha is estimated by using the Fung Hsieh (2004) seven-factor model. We form inception portfolios based on (1) the family structure of each inception (i.e., the stand-alone inception or family-affiliated inception including nonclone inceptions and clone inceptions), and (2) the strategy-based identification of each inception (i.e., cold or hot inception). In any month, inception portfolios are formed from new hedge fund inceptions of a given family structure and strategy identification over the prior three-month period. Each inception is then held in its corresponding portfolio for a 48-month or a 72-month holding period after its inception. The holding period follows the actual inception date of each fund. Within the holding period, we exclude backfilled returns and we require at least 12 monthly return observations for an inception to be included in any inception portfolio. Funds are equally weighted, rebalanced at the beginning of each month. The regression equation is presented in Table IV. t-statistics are in parentheses. Statistical significance at the 1%, 5%, and 10% levels is denoted by ***, **, and *, respectively.


Panel A1: Portfolio Alphas by Family Structure and Hot/Cold Strategy Identification (48-Month Holding Period)

Stand-Alone

(New Family) Family-Affiliated

Inceptions

Inceptions Nonclone Clone Cold Inceptions 0.682% 0.496% 0.215%

(7.795) (4.762) (1.740) Hot Inceptions 0.339% 0.165% 0.131%

(2.242) (1.335) (1.283) Cold minus Hot Spread 0.343% 0.331% 0.092%

(2.343) (2.274) (0.632)

Panel B1: Cold-Hot Corner Portfolio Spreads (48-Month Holding Period) Cold Nonclone minus Hot Clone Spread 0.365%

(2.945) Cold Stand-Alone minus Hot Clone Spread 0.550% (4.330)

Panel A2: Portfolio Alphas by Family Structure and Hot/Cold Strategy Identification (72-Month Holding Period)

Stand-Alone


Inceptions

Inceptions Nonclone Clone Cold Inceptions 0.568% 0.470% 0.265%

(7.388) (4.767) (2.385) Hot Inceptions 0.262% 0.073% 0.101%

(2.133) (0.652) (1.007) Cold minus Hot Spread 0.306% 0.397% 0.164%

(2.813) (3.290) (1.283) Panel B2: Cold-Hot Corner Portfolio Spreads (72-month holding period)

Cold Nonclone minus Hot Clone Spread 0.368%

(3.295) Cold Stand-Alone minus Hot Clone Spread 0.466% (4.064)


Table AVIII Strategy Demand and Family Structure: Evidence Using Alternative Measure of Investor Demand

This table presents a two-way summary of the risk-adjusted returns (alphas) of different types of inception portfolios using an alternative strategy to identify cold/hot inceptions. In this test, a strategy is classified as hot (cold) if its normalized inception is among the top (bottom) 30% of all strategies over the 36 months prior to inception. For each strategy, its inception is normalized by dividing by the number of funds in the strategy at the beginning of that period. Each portfolio’s alpha is estimated by using the Fung Hsieh (2004) seven-factor model. We form inception portfolios based on (1) the family structure of each inception (i.e., the stand-alone inception or family-affiliated inception including nonclone inceptions and clone inceptions), and (2) the strategy-based identification of each inception (i.e., cold or hot inception). In any month, inception portfolios are formed from new hedge fund inceptions of a given family structure and strategy identification over the prior three-month period. Each inception is then held in its corresponding portfolio for a 60-month holding period after its inception. The holding period follows the actual inception date of each fund. Within the holding period, we exclude backfilled returns and we require at least 12 monthly return observations for an inception to be included in any inception portfolio. Funds are equally weighted, rebalanced at the beginning of each month. The regression equation is presented in Table IV. t-statistics are in parentheses. Statistical significance at the 1%, 5%, and 10% levels is denoted by ***, **, and *, respectively.


Stand-Alone


Inceptions Inceptions Nonclone Clone

Cold Inceptions 0.587% 0.347% 0.317%

(8.467) (5.090) (3.752)

Hot Inceptions 0.346% 0.197% 0.171%

(4.098) (2.602) (2.290)

Cold minus Hot Spread 0.241% 0.150% 0.146%

(2.337) (1.904) (1.386)


Cold Nonclone minus Hot Clone Spread 0.189%

(2.117)

Cold Stand-Alone minus Hot Clone Spread 0.410%

(4.256)

Documents

The Economics of Hedge Fund Startups: Theory and Empirical